[{"data":1,"prerenderedAt":2246},["ShallowReactive",2],{"blog-post-en-\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700":3,"blog-posts-sidebar-en":1729},{"id":4,"title":5,"body":6,"categories":1710,"date":1715,"description":1716,"extension":1717,"heading":1718,"image":1719,"meta":1720,"navigation":973,"originalUrl":1721,"path":1722,"seo":1723,"slug":1724,"socialImage":1725,"stem":1726,"updated":1727,"__hash__":1728},"blog\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700.md","Qwen3.8 Flash Next in 4-bit on 2 × Radeon R9700: tested against 16-bit",{"type":7,"value":8,"toc":1695},"minimark",[9,34,57,65,70,84,88,94,162,171,188,193,202,205,210,286,295,345,366,387,396,400,413,416,421,459,484,488,509,512,521,583,634,655,659,678,681,686,733,756,789,793,814,827,838,855,872,876,914,923,1165,1195,1216,1219,1285,1301,1305,1319,1344,1348,1355,1359,1372,1391,1691],[10,11,12,13,17,18,21,22],"p",{},"The 16-bit Qwen3.8 Flash Next checkpoint is ",[14,15,16],"strong",{},"360 GB",". Two Radeon AI PRO R9700 cards offer 64 GB of VRAM between them, so the original model is far out of reach. Our new ",[14,19,20],{},"4-bit release runs on those two cards",", with results close to the 16-bit reference on math, Python coding and knowledge tasks.",[23,24,25],"sup",{},[26,27,33],"a",{"href":28,"ariaDescribedBy":29,"dataFootnoteRef":31,"id":32},"#user-content-fn-model",[30],"footnote-label","","user-content-fnref-model","1",[10,35,36,37,40,41,44,45,48,49],{},"The practical result: ",[14,38,39],{},"189.8 tokens\u002Fs for weighted single-request generation",", ",[14,42,43],{},"420.7 tokens\u002Fs across eight concurrent requests",", and an optional RAM prefix cache for long coding conversations. It is regular vLLM with the Paiton plugin and our own kernels. The benchmarked setup needs a ",[14,46,47],{},"256 GB RAM host",", because less-used experts and large tables live in system memory.",[23,50,51],{},[26,52,56],{"href":53,"ariaDescribedBy":54,"dataFootnoteRef":31,"id":55},"#user-content-fn-fourbit",[30],"user-content-fnref-fourbit","2",[10,58,59,60,64],{},"This follows our ",[26,61,63],{"href":62},"\u002Fblog\u002Fpaiton-qwen38-flash-next-radeon-ai-pro-r9700","3-bit Flash Next release",". Here the priority is retaining more of the original model's quality.",[66,67,69],"h2",{"id":68},"accuracy-against-the-16-bit-model","Accuracy against the 16-bit model",[10,71,72,73,77,78],{},"We ran both models on the same questions, with the same prompts, per-question sampling seeds and thinking-on chat template. The 4-bit model used the shipped ",[74,75,76],"code",{},"4bit"," serving mode; the 16-bit reference used the original checkpoint.",[23,79,80],{},[26,81,33],{"href":28,"ariaDescribedBy":82,"dataFootnoteRef":31,"id":83},[30],"user-content-fnref-model-2",[85,86],"flash-next-four-bit-figure",{"kind":87},"accuracy",[10,89,90],{},[91,92,93],"em",{},"Task accuracy; higher is better. GSM8K and HumanEval use their full test sets. MMLU-Pro uses a fixed 1,400-question subset, 100 questions in each of 14 categories. HumanEval is pass@1. Both models have a 16,384-token response limit.",[95,96,97,117],"table",{},[98,99,100],"thead",{},[101,102,103,107,111,114],"tr",{},[104,105,106],"th",{},"Task",[104,108,110],{"align":109},"right","Questions",[104,112,113],{"align":109},"16-bit",[104,115,116],{"align":109},"Paiton 4-bit",[118,119,120,135,148],"tbody",{},[101,121,122,126,129,132],{},[123,124,125],"td",{},"GSM8K, math",[123,127,128],{"align":109},"1,319",[123,130,131],{"align":109},"96.9%",[123,133,134],{"align":109},"96.8%",[101,136,137,140,143,146],{},[123,138,139],{},"HumanEval, Python code",[123,141,142],{"align":109},"164",[123,144,145],{"align":109},"96.3%",[123,147,145],{"align":109},[101,149,150,153,156,159],{},[123,151,152],{},"MMLU-Pro, knowledge",[123,154,155],{"align":109},"1,400",[123,157,158],{"align":109},"84.7%",[123,160,161],{"align":109},"83.7%",[10,163,164,165],{},"GSM8K differs by 0.1 percentage point; HumanEval is tied. In a second seed set, the 16-bit model itself changed by 0.0 points on GSM8K and 1.2 points on HumanEval. The paired tests for the two quantized comparisons both give p = 1.00.",[23,166,167],{},[26,168,33],{"href":28,"ariaDescribedBy":169,"dataFootnoteRef":31,"id":170},[30],"user-content-fnref-model-3",[10,172,173,174,177,178,181,182],{},"MMLU-Pro is one point lower for the 4-bit model, with a paired 95% interval of ",[14,175,176],{},"−2.3 to +0.3 points"," and p = 0.16. That does not establish statistical equivalence. More 4-bit replies reached the response limit: 104 versus 79. On the ",[14,179,180],{},"1,271 questions where neither model reached it",", scores are 91.1% for 4-bit and 91.0% for 16-bit. That subset helps describe the gap, but does not prove a longer limit would remove it.",[23,183,184],{},[26,185,33],{"href":28,"ariaDescribedBy":186,"dataFootnoteRef":31,"id":187},[30],"user-content-fnref-model-4",[189,190,192],"h3",{"id":191},"beyond-task-scores","Beyond task scores",[10,194,195,196],{},"Task scores can hide differences in the model's token predictions. We also measured perplexity, long-document negative log-likelihood and full-vocabulary KL divergence against BF16, the 16-bit reference.",[23,197,198],{},[26,199,33],{"href":28,"ariaDescribedBy":200,"dataFootnoteRef":31,"id":201},[30],"user-content-fnref-model-5",[85,203],{"kind":204},"quality",[10,206,207],{},[91,208,209],{},"Perplexity increase relative to BF16; lower is better. The two context lengths are evaluated separately. KL is a different measure: the average difference between next-token probability distributions on 170,884 text and assistant positions, with each model free to choose its experts. Zero KL would mean identical distributions.",[95,211,212,228],{},[98,213,214],{},[101,215,216,219,222,225],{},[104,217,218],{},"Text and context length",[104,220,221],{"align":109},"16-bit perplexity",[104,223,224],{"align":109},"4-bit perplexity",[104,226,227],{"align":109},"Increase",[118,229,230,244,258,272],{},[101,231,232,235,238,241],{},[123,233,234],{},"WikiText-2 test, 512 tokens",[123,236,237],{"align":109},"4.589",[123,239,240],{"align":109},"4.629",[123,242,243],{"align":109},"0.87%",[101,245,246,249,252,255],{},[123,247,248],{},"WikiText-2 test, 4,096 tokens",[123,250,251],{"align":109},"3.929",[123,253,254],{"align":109},"3.961",[123,256,257],{"align":109},"0.80%",[101,259,260,263,266,269],{},[123,261,262],{},"Unseen text, 512 tokens",[123,264,265],{"align":109},"6.938",[123,267,268],{"align":109},"6.954",[123,270,271],{"align":109},"0.23%",[101,273,274,277,280,283],{},[123,275,276],{},"Unseen text, 4,096 tokens",[123,278,279],{"align":109},"6.163",[123,281,282],{"align":109},"6.187",[123,284,285],{"align":109},"0.38%",[10,287,288,289],{},"The unseen text contains 22 documents, 262,144 tokens in total, never used to build the quant. Perplexity uses raw text without a chat template, scoring the second half of each context chunk.",[23,290,291],{},[26,292,33],{"href":28,"ariaDescribedBy":293,"dataFootnoteRef":31,"id":294},[30],"user-content-fnref-model-6",[95,296,297,310],{},[98,298,299],{},[101,300,301,304,307],{},[104,302,303],{},"Configuration",[104,305,306],{"align":109},"Routed-expert bits per weight",[104,308,309],{"align":109},"KL to BF16",[118,311,312,323,334],{},[101,313,314,317,320],{},[123,315,316],{},"Two correct BF16 implementations",[123,318,319],{"align":109},"16",[123,321,322],{"align":109},"0.0077",[101,324,325,328,331],{},[123,326,327],{},"Paiton 4-bit, INT6 trunk and FP8 cache",[123,329,330],{"align":109},"3.926",[123,332,333],{"align":109},"0.0306",[101,335,336,339,342],{},[123,337,338],{},"Standard MXFP4 reference, 8-bit trunk and BF16 cache",[123,340,341],{"align":109},"4.25",[123,343,344],{"align":109},"0.0515",[10,346,347,348,351,352,355,356,359,360],{},"The MXFP4 row is ",[14,349,350],{},"our own measurement on the same experts",", not another serving engine. The 4-bit served configuration has ",[14,353,354],{},"41% lower KL"," and about ",[14,357,358],{},"8% fewer bits per expert weight",", including scales. The trunk and cache also differ, so this does not isolate the weight format's effect or mean 41% better task accuracy.",[23,361,362],{},[26,363,33],{"href":28,"ariaDescribedBy":364,"dataFootnoteRef":31,"id":365},[30],"user-content-fnref-model-7",[10,367,368,369,372,373,376,377,380,381],{},"Across 44 held-out documents of 8K to 32K tokens, negative log-likelihood is ",[14,370,371],{},"within 0.16% of BF16",". A needle-retrieval test at ",[14,374,375],{},"131,072 tokens"," returns ",[14,378,379],{},"40 out of 40",", with all 419 answer tokens matching the reference.",[23,382,383],{},[26,384,33],{"href":28,"ariaDescribedBy":385,"dataFootnoteRef":31,"id":386},[30],"user-content-fnref-model-8",[10,388,389,390],{},"These distribution and retrieval checks emulate the default served arithmetic: INT6 trunk, int8 prompt activations and FP8 KV cache. The task benchmarks above run through the server. The separate 200K serving mode uses an 8-bit trunk and was not accuracy-tested separately.",[23,391,392],{},[26,393,33],{"href":28,"ariaDescribedBy":394,"dataFootnoteRef":31,"id":395},[30],"user-content-fnref-model-9",[66,397,399],{"id":398},"useful-speed-on-two-workstation-gpus","Useful speed on two workstation GPUs",[10,401,402,403,406,407],{},"The published performance measurements use ",[14,404,405],{},"BetterBench 0.6.0, standard profile",", through the exact release image and launcher, on two R9700 cards in a PCIe 4.0 x16 host with 251 GiB of RAM.",[23,408,409],{},[26,410,56],{"href":53,"ariaDescribedBy":411,"dataFootnoteRef":31,"id":412},[30],"user-content-fnref-fourbit-2",[85,414],{"kind":415},"concurrency",[10,417,418],{},[91,419,420],{},"Generated tokens per second; higher is better. The single-request number is a weighted decode score. The eight-request number is combined server throughput from a separate concurrency workload, not a rate per user or an eightfold scaling comparison.",[95,422,423,433],{},[98,424,425],{},[101,426,427,430],{},[104,428,429],{},"Measurement",[104,431,432],{"align":109},"Result",[118,434,435,443,451],{},[101,436,437,440],{},[123,438,439],{},"Weighted single-request decode, median of three runs",[123,441,442],{"align":109},"189.8 tokens\u002Fs",[101,444,445,448],{},[123,446,447],{},"Eight concurrent requests, aggregate",[123,449,450],{"align":109},"420.7 tokens\u002Fs",[101,452,453,456],{},[123,454,455],{},"Short-prompt time to first token, median",[123,457,458],{"align":109},"121 ms",[10,460,461,462,465,466,469,470,476],{},"All ",[14,463,464],{},"48 of 48 requests"," completed at concurrency eight. BetterBench measures eight decode categories, but the weighted score combines ",[14,467,468],{},"six",": code 30%, reasoning 20%, prose and JSON 15% each, file editing and summarization 10% each. Chat and math are measured without contributing to that score.",[23,471,472],{},[26,473,56],{"href":53,"ariaDescribedBy":474,"dataFootnoteRef":31,"id":475},[30],"user-content-fnref-fourbit-3",[23,477,478],{},[26,479,483],{"href":480,"ariaDescribedBy":481,"dataFootnoteRef":31,"id":482},"#user-content-fn-betterbench",[30],"user-content-fnref-betterbench","3",[189,485,487],{"id":486},"long-prompts","Long prompts",[10,489,490,491,494,495,498,499,502,503],{},"Prefill is reading the input before the model starts answering. The long-context mode processes the nominal 64K prompt depth at ",[14,492,493],{},"6,991 input tokens\u002Fs",". A separate cold-cache check with ",[14,496,497],{},"exactly 64,000 prompt tokens"," measures ",[14,500,501],{},"7,162 tokens\u002Fs",".",[23,504,505],{},[26,506,56],{"href":53,"ariaDescribedBy":507,"dataFootnoteRef":31,"id":508},[30],"user-content-fnref-fourbit-4",[85,510],{"kind":511},"prefill",[10,513,514],{},[91,515,516,517,520],{},"Input tokens processed per second; higher is better. BetterBench prompt-depth settings are nominal, not exact token counts. These rows use ",[74,518,519],{},"4bit-200k",", with a 200,000-token window and speculative decoding off. They are separate from the default decode measurements.",[95,522,523,533],{},[98,524,525],{},[101,526,527,530],{},[104,528,529],{},"Prompt-depth setting",[104,531,532],{"align":109},"Input tokens\u002Fs",[118,534,535,543,551,559,567,575],{},[101,536,537,540],{},[123,538,539],{},"2K",[123,541,542],{"align":109},"5,564",[101,544,545,548],{},[123,546,547],{},"8K",[123,549,550],{"align":109},"6,771",[101,552,553,556],{},[123,554,555],{},"16K",[123,557,558],{"align":109},"6,641",[101,560,561,564],{},[123,562,563],{},"32K",[123,565,566],{"align":109},"7,044",[101,568,569,572],{},[123,570,571],{},"64K",[123,573,574],{"align":109},"6,991",[101,576,577,580],{},[123,578,579],{},"128K",[123,581,582],{"align":109},"6,857",[95,584,585,601],{},[98,586,587],{},[101,588,589,592,595,598],{},[104,590,591],{},"Mode",[104,593,594],{"align":109},"Context window",[104,596,597],{},"Speculative decoding",[104,599,600],{"align":109},"Shared KV-cache capacity",[118,602,603,619],{},[101,604,605,610,613,616],{},[123,606,607,609],{},[74,608,76],{},", default",[123,611,612],{"align":109},"98,304 tokens",[123,614,615],{},"On, depth 3",[123,617,618],{"align":109},"305,338 tokens",[101,620,621,625,628,631],{},[123,622,623],{},[74,624,519],{},[123,626,627],{"align":109},"200,000 tokens",[123,629,630],{},"Off",[123,632,633],{"align":109},"247,058 tokens",[10,635,636,637,640,641,647],{},"Both use an ",[14,638,639],{},"FP8 E4M3 KV cache",". Cache capacity is shared across requests; it is not the maximum window of a single conversation. Each window must fit the prompt, chat formatting and generated answer together.",[23,642,643],{},[26,644,33],{"href":28,"ariaDescribedBy":645,"dataFootnoteRef":31,"id":646},[30],"user-content-fnref-model-10",[23,648,649],{},[26,650,654],{"href":651,"ariaDescribedBy":652,"dataFootnoteRef":31,"id":653},"#user-content-fn-modes",[30],"user-content-fnref-modes","4",[189,656,658],{"id":657},"faster-repeated-conversations-with-the-ram-prefix-cache","Faster repeated conversations with the RAM prefix cache",[10,660,661,662,668,669,502,672],{},"Coding agents resend their conversation on each turn. The optional ",[14,663,664,667],{},[74,665,666],{},"4bit-cache"," mode"," reuses unchanged prefixes in 2,048-token blocks and keeps evicted prefix blocks in an ",[14,670,671],{},"8 GiB pinned RAM tier",[23,673,674],{},[26,675,56],{"href":53,"ariaDescribedBy":676,"dataFootnoteRef":31,"id":677},[30],"user-content-fnref-fourbit-5",[85,679],{"kind":680},"prefix-cache",[10,682,683],{},[91,684,685],{},"Time to first token in separate repeated-prompt cache checks; lower is better. A cache hit requires a reusable prefix still present in the cache. These are not the cold-cache BetterBench prefill measurements.",[95,687,688,701],{},[98,689,690],{},[101,691,692,695,698],{},[104,693,694],{},"Repeated prompt",[104,696,697],{"align":109},"Cold first token",[104,699,700],{"align":109},"RAM-tier cache hit",[118,702,703,713,723],{},[101,704,705,707,710],{},[123,706,563],{},[123,708,709],{"align":109},"6.3 s",[123,711,712],{"align":109},"1.0 s",[101,714,715,717,720],{},[123,716,571],{},[123,718,719],{"align":109},"12.3 s",[123,721,722],{"align":109},"1.1 s",[101,724,725,728,731],{},[123,726,727],{},"90K",[123,729,730],{"align":109},"16.2 s",[123,732,722],{"align":109},[10,734,735,736,739,740,743,744,750],{},"The tested RAM-tier hits give ",[14,737,738],{},"identical output to a cold start",". The tradeoff is about ",[14,741,742],{},"22% less multi-user throughput"," in the cache-mode validation, 321.4 versus 411.3 aggregate tokens\u002Fs. That validation baseline is separate from the 420.7 BetterBench headline. New prompts also start 0.1 to 0.2 seconds later.",[23,745,746],{},[26,747,56],{"href":53,"ariaDescribedBy":748,"dataFootnoteRef":31,"id":749},[30],"user-content-fnref-fourbit-6",[23,751,752],{},[26,753,654],{"href":651,"ariaDescribedBy":754,"dataFootnoteRef":31,"id":755},[30],"user-content-fnref-modes-2",[10,757,758,759,762,763,765,766,768,769,771,772,502,775,781],{},"The full RAM tier needs about ",[14,760,761],{},"108 GiB of free GTT",", including its extra pinned memory. With less available, the launcher warns and shrinks or disables the RAM tier, keeping VRAM prefix caching. Use ",[74,764,666],{}," explicitly; the ordinary ",[74,767,76],{}," and ",[74,770,519],{}," modes reject ",[74,773,774],{},"--prefix-caching",[23,776,777],{},[26,778,56],{"href":53,"ariaDescribedBy":779,"dataFootnoteRef":31,"id":780},[30],"user-content-fnref-fourbit-7",[23,782,783],{},[26,784,788],{"href":785,"ariaDescribedBy":786,"dataFootnoteRef":31,"id":787},"#user-content-fn-launcher",[30],"user-content-fnref-launcher","5",[66,790,792],{"id":791},"what-4-bit-means-here","What 4-bit means here",[10,794,795,796,799,800,803,804,807,808],{},"Flash Next is a mixture-of-experts model. In each layer, the ",[14,797,798],{},"410 most-used experts use INT4",", calibrated with GPTQ in a rotated basis. The ",[14,801,802],{},"102 least-used experts use the 3-bit codebook format"," from our earlier release. Together, the routed experts average ",[14,805,806],{},"3.926 bits per weight",", including scales. That average belongs to the mixed expert set, not to the 3-bit format alone.",[23,809,810],{},[26,811,33],{"href":28,"ariaDescribedBy":812,"dataFootnoteRef":31,"id":813},[30],"user-content-fnref-model-11",[10,815,816,817,820,821],{},"The default serving mode uses a ",[14,818,819],{},"6-bit trunk",", while other parts retain 8-bit or higher precision. The MTP layer supplies speculative drafts, which the target model verifies. The model is not uniformly four bits.",[23,822,823],{},[26,824,33],{"href":28,"ariaDescribedBy":825,"dataFootnoteRef":31,"id":826},[30],"user-content-fnref-model-12",[10,828,829,830,833,834,837],{},"Getting here took roughly ",[14,831,832],{},"250 GPU-hours on an AMD Instinct MI355X system"," for quantization, tuning and accuracy testing, plus about ",[14,835,836],{},"100 hours on the two R9700 cards"," testing the served model. Those estimates cover our 3-bit and 4-bit Flash Next releases together.",[10,839,840,841,844,845,848,849],{},"The full 4-bit expert set does not fit next to the runtime and cache on two 32 GB cards. About ",[14,842,843],{},"20.8 GiB of experts per card stay in VRAM","; the GPUs read the rest from pinned system memory when needed. The launcher pins roughly ",[14,846,847],{},"96 GiB in total",", about 55 GiB for tables and 41 GiB for expert copies. Host memory bandwidth and PCIe links therefore matter.",[23,850,851],{},[26,852,56],{"href":53,"ariaDescribedBy":853,"dataFootnoteRef":31,"id":854},[30],"user-content-fnref-fourbit-8",[10,856,857,858,861,862,865,866],{},"The benchmarked allocation uses GPU-mapped ",[14,859,860],{},"GTT memory",", with about ",[14,863,864],{},"98 GiB free GTT needed at baseline",". That is why we specify a 256 GB host. With less GTT, the launcher can fall back to ROCm user-pointer memory, warning about lower speed and possible pauses during host-memory compaction. It does not change kernel memory limits.",[23,867,868],{},[26,869,56],{"href":53,"ariaDescribedBy":870,"dataFootnoteRef":31,"id":871},[30],"user-content-fnref-fourbit-9",[66,873,875],{"id":874},"run-the-release","Run the release",[10,877,878,879,882,883,886,887,768,890,893,894,897,898,901,902,905,906],{},"Use Linux x86-64, ",[14,880,881],{},"two Radeon AI PRO R9700 cards",", a ",[14,884,885],{},"ROCm 10 host driver",", Docker with access to ",[74,888,889],{},"\u002Fdev\u002Fkfd",[74,891,892],{},"\u002Fdev\u002Fdri",", and Python ",[14,895,896],{},"3.10 or newer"," with ",[74,899,900],{},"huggingface_hub"," installed. Allow ",[14,903,904],{},"133 GB for weights",", plus the roughly 41 GB runtime image and working space.",[23,907,908],{},[26,909,913],{"href":910,"ariaDescribedBy":911,"dataFootnoteRef":31,"id":912},"#user-content-fn-release",[30],"user-content-fnref-release","6",[10,915,916,917],{},"The commands pin the launcher and download the complete model revision first. This avoids a first-start issue in the pinned launcher, which checks for the drafter before its automatic download. Use a fresh checkout if you already have local plugin changes.",[23,918,919],{},[26,920,788],{"href":785,"ariaDescribedBy":921,"dataFootnoteRef":31,"id":922},[30],"user-content-fnref-launcher-2",[924,925,929],"pre",{"className":926,"code":927,"language":928,"meta":31,"style":31},"language-bash shiki shiki-themes github-light github-dark","git clone https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin.git\ncd paiton-vllm-plugin\ngit checkout a985fa598d9a3c7536720304830a0e427f3af1f2\n\npython3 -m venv .venv\n. .venv\u002Fbin\u002Factivate\npython3 -m pip install huggingface_hub\npython3 - \u003C\u003C'PY'\nfrom huggingface_hub import snapshot_download\nsnapshot_download(\n    repo_id=\"EliovpAI\u002FQwen3.8-Flash-Next-W4A8-Paiton-RDNA4\",\n    revision=\"9ec8e306503d5f9c9e8fd9c0d77de8fa63428888\",\n    local_dir=\"flashnext-w4a8\",\n)\nPY\n(cd flashnext-w4a8 && sha256sum -c SHA256SUMS)\n\npython3 models\u002FQwen3.8-Flash-Next\u002Flaunch-flashnext.py \\\n  --mode 4bit --weights \"$PWD\u002Fflashnext-w4a8\" --dry-run\nbash models\u002FQwen3.8-Flash-Next\u002Frun-flashnext-4bit.sh \\\n  --weights \"$PWD\u002Fflashnext-w4a8\" --host 127.0.0.1\n","bash",[74,930,931,947,957,968,975,990,998,1014,1029,1035,1041,1047,1053,1059,1065,1071,1097,1102,1113,1137,1147],{"__ignoreMap":31},[932,933,936,940,944],"span",{"class":934,"line":935},"line",1,[932,937,939],{"class":938},"sScJk","git",[932,941,943],{"class":942},"sZZnC"," clone",[932,945,946],{"class":942}," https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin.git\n",[932,948,950,954],{"class":934,"line":949},2,[932,951,953],{"class":952},"sj4cs","cd",[932,955,956],{"class":942}," paiton-vllm-plugin\n",[932,958,960,962,965],{"class":934,"line":959},3,[932,961,939],{"class":938},[932,963,964],{"class":942}," checkout",[932,966,967],{"class":942}," a985fa598d9a3c7536720304830a0e427f3af1f2\n",[932,969,971],{"class":934,"line":970},4,[932,972,974],{"emptyLinePlaceholder":973},true,"\n",[932,976,978,981,984,987],{"class":934,"line":977},5,[932,979,980],{"class":938},"python3",[932,982,983],{"class":952}," -m",[932,985,986],{"class":942}," venv",[932,988,989],{"class":942}," .venv\n",[932,991,993,995],{"class":934,"line":992},6,[932,994,502],{"class":952},[932,996,997],{"class":942}," .venv\u002Fbin\u002Factivate\n",[932,999,1001,1003,1005,1008,1011],{"class":934,"line":1000},7,[932,1002,980],{"class":938},[932,1004,983],{"class":952},[932,1006,1007],{"class":942}," pip",[932,1009,1010],{"class":942}," install",[932,1012,1013],{"class":942}," huggingface_hub\n",[932,1015,1017,1019,1022,1026],{"class":934,"line":1016},8,[932,1018,980],{"class":938},[932,1020,1021],{"class":942}," -",[932,1023,1025],{"class":1024},"szBVR"," \u003C\u003C",[932,1027,1028],{"class":942},"'PY'\n",[932,1030,1032],{"class":934,"line":1031},9,[932,1033,1034],{"class":942},"from huggingface_hub import snapshot_download\n",[932,1036,1038],{"class":934,"line":1037},10,[932,1039,1040],{"class":942},"snapshot_download(\n",[932,1042,1044],{"class":934,"line":1043},11,[932,1045,1046],{"class":942},"    repo_id=\"EliovpAI\u002FQwen3.8-Flash-Next-W4A8-Paiton-RDNA4\",\n",[932,1048,1050],{"class":934,"line":1049},12,[932,1051,1052],{"class":942},"    revision=\"9ec8e306503d5f9c9e8fd9c0d77de8fa63428888\",\n",[932,1054,1056],{"class":934,"line":1055},13,[932,1057,1058],{"class":942},"    local_dir=\"flashnext-w4a8\",\n",[932,1060,1062],{"class":934,"line":1061},14,[932,1063,1064],{"class":942},")\n",[932,1066,1068],{"class":934,"line":1067},15,[932,1069,1070],{"class":942},"PY\n",[932,1072,1074,1078,1080,1083,1086,1089,1092,1095],{"class":934,"line":1073},16,[932,1075,1077],{"class":1076},"sVt8B","(",[932,1079,953],{"class":952},[932,1081,1082],{"class":942}," flashnext-w4a8",[932,1084,1085],{"class":1076}," && ",[932,1087,1088],{"class":938},"sha256sum",[932,1090,1091],{"class":952}," -c",[932,1093,1094],{"class":942}," SHA256SUMS",[932,1096,1064],{"class":1076},[932,1098,1100],{"class":934,"line":1099},17,[932,1101,974],{"emptyLinePlaceholder":973},[932,1103,1105,1107,1110],{"class":934,"line":1104},18,[932,1106,980],{"class":938},[932,1108,1109],{"class":942}," models\u002FQwen3.8-Flash-Next\u002Flaunch-flashnext.py",[932,1111,1112],{"class":952}," \\\n",[932,1114,1116,1119,1122,1125,1128,1131,1134],{"class":934,"line":1115},19,[932,1117,1118],{"class":952},"  --mode",[932,1120,1121],{"class":942}," 4bit",[932,1123,1124],{"class":952}," --weights",[932,1126,1127],{"class":942}," \"",[932,1129,1130],{"class":1076},"$PWD",[932,1132,1133],{"class":942},"\u002Fflashnext-w4a8\"",[932,1135,1136],{"class":952}," --dry-run\n",[932,1138,1140,1142,1145],{"class":934,"line":1139},20,[932,1141,928],{"class":938},[932,1143,1144],{"class":942}," models\u002FQwen3.8-Flash-Next\u002Frun-flashnext-4bit.sh",[932,1146,1112],{"class":952},[932,1148,1150,1153,1155,1157,1159,1162],{"class":934,"line":1149},21,[932,1151,1152],{"class":952},"  --weights",[932,1154,1127],{"class":942},[932,1156,1130],{"class":1076},[932,1158,1133],{"class":942},[932,1160,1161],{"class":952}," --host",[932,1163,1164],{"class":952}," 127.0.0.1\n",[10,1166,1167,1168,1171,1172,1175,1176,1179,1180,502,1183,1189],{},"The dry run reports RAM, GTT limits and GPU-link checks. The launcher pulls the runtime by digest. After the weight download, first start takes around ",[14,1169,1170],{},"five to six minutes","; wait for ",[74,1173,1174],{},"launcher: READY on port …",". The API is ",[74,1177,1178],{},"http:\u002F\u002F127.0.0.1:18982\u002Fv1",", with model name ",[74,1181,1182],{},"Qwen3.8-Flash-Next",[23,1184,1185],{},[26,1186,913],{"href":910,"ariaDescribedBy":1187,"dataFootnoteRef":31,"id":1188},[30],"user-content-fnref-release-2",[23,1190,1191],{},[26,1192,654],{"href":651,"ariaDescribedBy":1193,"dataFootnoteRef":31,"id":1194},[30],"user-content-fnref-modes-3",[10,1196,1197,1198,1201,1202,1205,1206,1209,1210],{},"The default server binds ",[14,1199,1200],{},"all network interfaces without an API key",". The example restricts it to localhost. For trusted remote clients, protect access with a firewall or set ",[74,1203,1204],{},"PAITON_API_KEY"," in the environment. Prefer the environment variable to ",[74,1207,1208],{},"--api-key",", which exposes the key in the process list.",[23,1211,1212],{},[26,1213,913],{"href":910,"ariaDescribedBy":1214,"dataFootnoteRef":31,"id":1215},[30],"user-content-fnref-release-3",[10,1217,1218],{},"Stop the running server before choosing another mode:",[924,1220,1222],{"className":926,"code":1221,"language":928,"meta":31,"style":31},"# 200K window, speculative decoding off\nbash models\u002FQwen3.8-Flash-Next\u002Frun-flashnext-4bit-200k.sh \\\n  --weights \"$PWD\u002Fflashnext-w4a8\" --host 127.0.0.1\n\n# Repeated coding conversations, opt-in RAM prefix cache\nbash models\u002FQwen3.8-Flash-Next\u002Frun-flashnext-4bit-cache.sh \\\n  --weights \"$PWD\u002Fflashnext-w4a8\" --host 127.0.0.1\n",[74,1223,1224,1230,1239,1253,1257,1262,1271],{"__ignoreMap":31},[932,1225,1226],{"class":934,"line":935},[932,1227,1229],{"class":1228},"sJ8bj","# 200K window, speculative decoding off\n",[932,1231,1232,1234,1237],{"class":934,"line":949},[932,1233,928],{"class":938},[932,1235,1236],{"class":942}," models\u002FQwen3.8-Flash-Next\u002Frun-flashnext-4bit-200k.sh",[932,1238,1112],{"class":952},[932,1240,1241,1243,1245,1247,1249,1251],{"class":934,"line":959},[932,1242,1152],{"class":952},[932,1244,1127],{"class":942},[932,1246,1130],{"class":1076},[932,1248,1133],{"class":942},[932,1250,1161],{"class":952},[932,1252,1164],{"class":952},[932,1254,1255],{"class":934,"line":970},[932,1256,974],{"emptyLinePlaceholder":973},[932,1258,1259],{"class":934,"line":977},[932,1260,1261],{"class":1228},"# Repeated coding conversations, opt-in RAM prefix cache\n",[932,1263,1264,1266,1269],{"class":934,"line":992},[932,1265,928],{"class":938},[932,1267,1268],{"class":942}," models\u002FQwen3.8-Flash-Next\u002Frun-flashnext-4bit-cache.sh",[932,1270,1112],{"class":952},[932,1272,1273,1275,1277,1279,1281,1283],{"class":934,"line":1000},[932,1274,1152],{"class":952},[932,1276,1127],{"class":942},[932,1278,1130],{"class":1076},[932,1280,1133],{"class":942},[932,1282,1161],{"class":952},[932,1284,1164],{"class":952},[10,1286,1287,1288,1294,1295],{},"The ",[26,1289,1293],{"href":1290,"rel":1291},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002Fa985fa598d9a3c7536720304830a0e427f3af1f2\u002Fmodels\u002FQwen3.8-Flash-Next\u002FREADME-4bit.md",[1292],"nofollow","4-bit setup guide"," documents the options. These weights require the Paiton runtime and plugin; stock model loaders do not support the format.",[23,1296,1297],{},[26,1298,33],{"href":28,"ariaDescribedBy":1299,"dataFootnoteRef":31,"id":1300},[30],"user-content-fnref-model-13",[66,1302,1304],{"id":1303},"also-in-this-release","Also in this release",[10,1306,1287,1307,1310,1311],{},[14,1308,1309],{},"3-bit Flash Next launcher"," gets a steadier startup on busy hosts, GPU-link and host-memory checks, and explicit 2,048-token prefix-cache blocks. The default all-interface binding and optional API key apply to both releases. The 3-bit weights and runtime image are unchanged.",[23,1312,1313],{},[26,1314,1318],{"href":1315,"ariaDescribedBy":1316,"dataFootnoteRef":31,"id":1317},"#user-content-fn-notes",[30],"user-content-fnref-notes","7",[10,1320,1321,1324,1325,1331,1332,1335,1336],{},[14,1322,1323],{},"Qwen3.8 27B on one R9700"," also gets a launcher update. Its ",[14,1326,1327,1328],{},"3-bit ",[74,1329,1330],{},"--mode long"," now holds ",[14,1333,1334],{},"316,325 tokens of shared cache",", and startup memory peaks are lower across the 3-bit modes. The MXFP4 configurations are unchanged.",[23,1337,1338],{},[26,1339,1343],{"href":1340,"ariaDescribedBy":1341,"dataFootnoteRef":31,"id":1342},"#user-content-fn-twentyseven",[30],"user-content-fnref-twentyseven","8",[66,1345,1347],{"id":1346},"working-toward-128-gb-hosts","Working toward 128 GB hosts",[10,1349,1350,1351,1354],{},"Not everyone has 256 GB of RAM. We are working on a 4-bit configuration that keeps the large n-gram table in ordinary, unpinned memory, with the goal of preserving bit-identical output. The target is ",[14,1352,1353],{},"128 GB machines",". This is work in progress, not a released or qualified configuration.",[66,1356,1358],{"id":1357},"credits-and-licensing","Credits and licensing",[10,1360,1361,1362,1365,1366,1371],{},"Qwen3.8 Flash Next is by the Qwen team; vLLM supplies the serving framework and BetterBench the performance suite. The model weights remain under ",[14,1363,1364],{},"Qwen Community License 1.0",", with commercial-use conditions. Review the ",[26,1367,1370],{"href":1368,"rel":1369},"https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3.8-Flash-Next\u002Fblob\u002Fde4b8e4d43b917e7706784d8bb445c9af86a3540\u002FLICENSE",[1292],"upstream license"," before deployment.",[10,1373,1374,1375,1380,1381,1385,1386,1390],{},"Read the ",[26,1376,1379],{"href":1377,"rel":1378},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FQwen3.8-Flash-Next-W4A8-Paiton-RDNA4\u002Fblob\u002F9ec8e306503d5f9c9e8fd9c0d77de8fa63428888\u002FREADME.md",[1292],"4-bit model card and accuracy details",", explore ",[26,1382,1384],{"href":1383},"\u002Fproducts\u002Fpaiton","Paiton",", or ",[26,1387,1389],{"href":1388},"\u002Fcontact","contact us"," about a local AMD AI workload.",[1392,1393,1396,1401],"section",{"className":1394,"dataFootnotes":31},[1395],"footnotes",[66,1397,1400],{"className":1398,"id":30},[1399],"sr-only","Footnotes",[1402,1403,1404,1510,1578,1591,1618,1638,1665,1678],"ol",{},[1405,1406,1408,1412,1413,1420,1421,1420,1428,1420,1435,1420,1442,1420,1449,1420,1456,1420,1463,1420,1470,1420,1478,1420,1486,1420,1494,1420,1502],"li",{"id":1407},"user-content-fn-model",[26,1409,1411],{"href":1377,"rel":1410},[1292],"W4A8 model card and evaluation methods, pinned revision",". ",[26,1414,1419],{"href":1415,"ariaLabel":1416,"className":1417,"dataFootnoteBackref":31},"#user-content-fnref-model","Back to reference 1",[1418],"data-footnote-backref","↩"," ",[26,1422,1419,1426],{"href":1423,"ariaLabel":1424,"className":1425,"dataFootnoteBackref":31},"#user-content-fnref-model-2","Back to reference 1-2",[1418],[23,1427,56],{},[26,1429,1419,1433],{"href":1430,"ariaLabel":1431,"className":1432,"dataFootnoteBackref":31},"#user-content-fnref-model-3","Back to reference 1-3",[1418],[23,1434,483],{},[26,1436,1419,1440],{"href":1437,"ariaLabel":1438,"className":1439,"dataFootnoteBackref":31},"#user-content-fnref-model-4","Back to reference 1-4",[1418],[23,1441,654],{},[26,1443,1419,1447],{"href":1444,"ariaLabel":1445,"className":1446,"dataFootnoteBackref":31},"#user-content-fnref-model-5","Back to reference 1-5",[1418],[23,1448,788],{},[26,1450,1419,1454],{"href":1451,"ariaLabel":1452,"className":1453,"dataFootnoteBackref":31},"#user-content-fnref-model-6","Back to reference 1-6",[1418],[23,1455,913],{},[26,1457,1419,1461],{"href":1458,"ariaLabel":1459,"className":1460,"dataFootnoteBackref":31},"#user-content-fnref-model-7","Back to reference 1-7",[1418],[23,1462,1318],{},[26,1464,1419,1468],{"href":1465,"ariaLabel":1466,"className":1467,"dataFootnoteBackref":31},"#user-content-fnref-model-8","Back to reference 1-8",[1418],[23,1469,1343],{},[26,1471,1419,1475],{"href":1472,"ariaLabel":1473,"className":1474,"dataFootnoteBackref":31},"#user-content-fnref-model-9","Back to reference 1-9",[1418],[23,1476,1477],{},"9",[26,1479,1419,1483],{"href":1480,"ariaLabel":1481,"className":1482,"dataFootnoteBackref":31},"#user-content-fnref-model-10","Back to reference 1-10",[1418],[23,1484,1485],{},"10",[26,1487,1419,1491],{"href":1488,"ariaLabel":1489,"className":1490,"dataFootnoteBackref":31},"#user-content-fnref-model-11","Back to reference 1-11",[1418],[23,1492,1493],{},"11",[26,1495,1419,1499],{"href":1496,"ariaLabel":1497,"className":1498,"dataFootnoteBackref":31},"#user-content-fnref-model-12","Back to reference 1-12",[1418],[23,1500,1501],{},"12",[26,1503,1419,1507],{"href":1504,"ariaLabel":1505,"className":1506,"dataFootnoteBackref":31},"#user-content-fnref-model-13","Back to reference 1-13",[1418],[23,1508,1509],{},"13",[1405,1511,1513,1412,1517,1420,1522,1420,1529,1420,1536,1420,1543,1420,1550,1420,1557,1420,1564,1420,1571],{"id":1512},"user-content-fn-fourbit",[26,1514,1516],{"href":1290,"rel":1515},[1292],"4-bit accuracy, speed and hardware guide, pinned release",[26,1518,1419],{"href":1519,"ariaLabel":1520,"className":1521,"dataFootnoteBackref":31},"#user-content-fnref-fourbit","Back to reference 2",[1418],[26,1523,1419,1527],{"href":1524,"ariaLabel":1525,"className":1526,"dataFootnoteBackref":31},"#user-content-fnref-fourbit-2","Back to reference 2-2",[1418],[23,1528,56],{},[26,1530,1419,1534],{"href":1531,"ariaLabel":1532,"className":1533,"dataFootnoteBackref":31},"#user-content-fnref-fourbit-3","Back to reference 2-3",[1418],[23,1535,483],{},[26,1537,1419,1541],{"href":1538,"ariaLabel":1539,"className":1540,"dataFootnoteBackref":31},"#user-content-fnref-fourbit-4","Back to reference 2-4",[1418],[23,1542,654],{},[26,1544,1419,1548],{"href":1545,"ariaLabel":1546,"className":1547,"dataFootnoteBackref":31},"#user-content-fnref-fourbit-5","Back to reference 2-5",[1418],[23,1549,788],{},[26,1551,1419,1555],{"href":1552,"ariaLabel":1553,"className":1554,"dataFootnoteBackref":31},"#user-content-fnref-fourbit-6","Back to reference 2-6",[1418],[23,1556,913],{},[26,1558,1419,1562],{"href":1559,"ariaLabel":1560,"className":1561,"dataFootnoteBackref":31},"#user-content-fnref-fourbit-7","Back to reference 2-7",[1418],[23,1563,1318],{},[26,1565,1419,1569],{"href":1566,"ariaLabel":1567,"className":1568,"dataFootnoteBackref":31},"#user-content-fnref-fourbit-8","Back to reference 2-8",[1418],[23,1570,1343],{},[26,1572,1419,1576],{"href":1573,"ariaLabel":1574,"className":1575,"dataFootnoteBackref":31},"#user-content-fnref-fourbit-9","Back to reference 2-9",[1418],[23,1577,1477],{},[1405,1579,1581,1412,1586],{"id":1580},"user-content-fn-betterbench",[26,1582,1585],{"href":1583,"rel":1584},"https:\u002F\u002Fgithub.com\u002FGGZ14\u002FBetterBench\u002Fblob\u002Fd00ad5ec8098c06584a88ec3468bacd37d5ed098\u002Fconfig\u002Fdefault.json",[1292],"BetterBench 0.6.0 workload weights",[26,1587,1419],{"href":1588,"ariaLabel":1589,"className":1590,"dataFootnoteBackref":31},"#user-content-fnref-betterbench","Back to reference 3",[1418],[1405,1592,1594,1412,1599,1420,1604,1420,1611],{"id":1593},"user-content-fn-modes",[26,1595,1598],{"href":1596,"rel":1597},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002Fa985fa598d9a3c7536720304830a0e427f3af1f2\u002Fmodels\u002FQwen3.8-Flash-Next\u002Fmodes.json",[1292],"Pinned weights, runtime digest and serving modes",[26,1600,1419],{"href":1601,"ariaLabel":1602,"className":1603,"dataFootnoteBackref":31},"#user-content-fnref-modes","Back to reference 4",[1418],[26,1605,1419,1609],{"href":1606,"ariaLabel":1607,"className":1608,"dataFootnoteBackref":31},"#user-content-fnref-modes-2","Back to reference 4-2",[1418],[23,1610,56],{},[26,1612,1419,1616],{"href":1613,"ariaLabel":1614,"className":1615,"dataFootnoteBackref":31},"#user-content-fnref-modes-3","Back to reference 4-3",[1418],[23,1617,483],{},[1405,1619,1621,1412,1626,1420,1631],{"id":1620},"user-content-fn-launcher",[26,1622,1625],{"href":1623,"rel":1624},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002Fa985fa598d9a3c7536720304830a0e427f3af1f2\u002Fmodels\u002FQwen3.8-Flash-Next\u002Flaunch-flashnext.py",[1292],"Public launcher preflight and RAM-tier sizing, pinned release",[26,1627,1419],{"href":1628,"ariaLabel":1629,"className":1630,"dataFootnoteBackref":31},"#user-content-fnref-launcher","Back to reference 5",[1418],[26,1632,1419,1636],{"href":1633,"ariaLabel":1634,"className":1635,"dataFootnoteBackref":31},"#user-content-fnref-launcher-2","Back to reference 5-2",[1418],[23,1637,56],{},[1405,1639,1641,1412,1646,1420,1651,1420,1658],{"id":1640},"user-content-fn-release",[26,1642,1645],{"href":1643,"rel":1644},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002Fa985fa598d9a3c7536720304830a0e427f3af1f2\u002Fmodels\u002FQwen3.8-Flash-Next\u002FREADME.md",[1292],"Launcher prerequisites and network configuration, pinned release",[26,1647,1419],{"href":1648,"ariaLabel":1649,"className":1650,"dataFootnoteBackref":31},"#user-content-fnref-release","Back to reference 6",[1418],[26,1652,1419,1656],{"href":1653,"ariaLabel":1654,"className":1655,"dataFootnoteBackref":31},"#user-content-fnref-release-2","Back to reference 6-2",[1418],[23,1657,56],{},[26,1659,1419,1663],{"href":1660,"ariaLabel":1661,"className":1662,"dataFootnoteBackref":31},"#user-content-fnref-release-3","Back to reference 6-3",[1418],[23,1664,483],{},[1405,1666,1668,1412,1673],{"id":1667},"user-content-fn-notes",[26,1669,1672],{"href":1670,"rel":1671},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002Fa985fa598d9a3c7536720304830a0e427f3af1f2\u002Fmodels\u002FQwen3.8-Flash-Next\u002FRELEASE-NOTES.md",[1292],"Flash Next release notes, pinned release",[26,1674,1419],{"href":1675,"ariaLabel":1676,"className":1677,"dataFootnoteBackref":31},"#user-content-fnref-notes","Back to reference 7",[1418],[1405,1679,1681,1412,1686],{"id":1680},"user-content-fn-twentyseven",[26,1682,1685],{"href":1683,"rel":1684},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002Fa985fa598d9a3c7536720304830a0e427f3af1f2\u002Fmodels\u002FQwen3.8-MXFP4-DFlash2\u002FREFERENCE.md#launcher-update-11-october-2026",[1292],"27B 3-bit launcher update of 11 October, pinned release",[26,1687,1419],{"href":1688,"ariaLabel":1689,"className":1690,"dataFootnoteBackref":31},"#user-content-fnref-twentyseven","Back to reference 8",[1418],[1692,1693,1694],"style",{},"html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}",{"title":31,"searchDepth":949,"depth":949,"links":1696},[1697,1700,1704,1705,1706,1707,1708,1709],{"id":68,"depth":949,"text":69,"children":1698},[1699],{"id":191,"depth":959,"text":192},{"id":398,"depth":949,"text":399,"children":1701},[1702,1703],{"id":486,"depth":959,"text":487},{"id":657,"depth":959,"text":658},{"id":791,"depth":949,"text":792},{"id":874,"depth":949,"text":875},{"id":1303,"depth":949,"text":1304},{"id":1346,"depth":949,"text":1347},{"id":1357,"depth":949,"text":1358},{"id":30,"depth":949,"text":1400},[1384,1711,1712,1713,1714],"AMD Radeon","Local AI","Qwen","Quantization","2026-10-12T15:00:00Z","Paiton runs Qwen3.8 Flash Next in 4-bit on two R9700 cards: 96.8% GSM8K, 96.3% HumanEval, 189.8 tokens\u002Fs and an optional RAM prefix cache. Accuracy, speed and setup.","md","4-bit on two cards, tested against 16-bit.","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit\u002Fhero.webp",{},"https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700","\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700",{"title":5,"description":1716},"paiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit\u002Fsocial.webp","blog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700",null,"4JwkGwHWsVLVwL2O_sC8UHpzPed0Ki_CciJ2ogBR6jk",[1730,1732,1740,1749,1759,1772,1782,1794,1803,1818,1827,1841,1873,1885,1907,1925,1944,1962,1980,1997,2009,2025,2040,2052,2061,2069,2084,2096,2107,2118,2128,2141,2151,2164,2175,2185,2196,2205,2217,2228,2237],{"path":1722,"title":5,"description":1716,"date":1715,"slug":1724,"image":1719,"originalUrl":1721,"categories":1731},[1384,1711,1712,1713,1714],{"path":62,"title":1733,"description":1734,"date":1735,"slug":1736,"image":1737,"originalUrl":1738,"categories":1739},"Qwen3.8 Flash Next on 2 × Radeon AI PRO R9700: 216 tokens\u002Fs, 200K context","Paiton runs Qwen3.8 Flash Next with 3-bit experts on two R9700 cards: 216.3 tokens\u002Fs, 565.1 aggregate tokens\u002Fs and a separate 200K mode. Benchmarks, quality and setup.","2026-10-10T10:30:00Z","paiton-qwen38-flash-next-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-flash-next\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-radeon-ai-pro-r9700",[1384,1711,1712,1713,1714],{"path":1741,"title":1742,"description":1743,"date":1744,"slug":1745,"image":1746,"originalUrl":1747,"categories":1748},"\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700","Qwen3.8 27B on 1 × Radeon AI PRO R9700: 3-bit weights, 20% faster decode","Qwen3.8 27B on one R9700: 19.9% faster weighted decode, a new 4-bit cache, 200K context and optional vision. Benchmarks and current Paiton setup.","2026-09-27T09:00:00Z","paiton-qwen38-w3a4-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-w3a4\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700",[1384,1711,1712,1713,1714],{"path":1750,"title":1751,"description":1752,"date":1753,"slug":1754,"image":1755,"originalUrl":1756,"categories":1757},"\u002Fblog\u002Fpaiton-qwen-image-21-radeon-ai-pro-r9700","Qwen-Image 2.1 on 1 × Radeon AI PRO R9700: 2048×2048 images in 103 seconds","Generate 2048×2048 Qwen-Image 2.1 images locally with 1 × Radeon AI PRO R9700. The released Paiton v1.0.2 container measured 103.29 seconds per warm request, through PNG delivery.","2026-09-23T09:00:00Z","paiton-qwen-image-21-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen-image-21\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen-image-21-radeon-ai-pro-r9700",[1384,1711,1712,1758,1713],"Image Generation",{"path":1760,"title":1761,"description":1762,"date":1763,"slug":1764,"image":1765,"originalUrl":1766,"categories":1767},"\u002Fblog\u002Fpaiton-qwen38-mxfp4-dflash2-r9700","Qwen3.8: 400.7 tok\u002Fs on R9700 | Paiton","Qwen3.8 on one R9700: 400.7 aggregate tok\u002Fs with ROCm 10 and vLLM 0.29, plus public 200K\u002F220K chat profiles. Benchmarks, limits and launch commands.","2026-09-16T07:30:00Z","paiton-qwen38-mxfp4-dflash2-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-mxfp4\u002Fupdate-2026-09-19\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-mxfp4-dflash2-r9700",[1384,1711,1768,1769,1770,1771],"vLLM","Qwen3.8","Inference Optimization","DFlash2",{"path":1773,"title":1774,"description":1775,"date":1776,"slug":1777,"image":1778,"originalUrl":1779,"categories":1780},"\u002Fblog\u002Fpaiton-qwen38-neo-gguf-vllm-r9700","Qwen3.8 GGUF in vLLM: Faster Responses on One Radeon","Run the original NEO CODER MAX GGUF in vLLM with Paiton on an R9700. Explore measured latency gains, image input and local deployment.","2026-09-14T07:30:00Z","paiton-qwen38-neo-gguf-vllm-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-neo-gguf\u002F00-hero-neo-gguf-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-neo-gguf-vllm-r9700",[1384,1711,1712,1781,1768],"GGUF",{"path":1783,"title":1784,"description":1785,"date":1786,"slug":1787,"image":1788,"originalUrl":1789,"categories":1790},"\u002Fblog\u002Fpaiton-minimax-h3-radeon-ai-pro-r9700","MiniMax H3 on Radeon: 15-Second Video With Native Sound","Paiton generates a 15-second MiniMax H3 video with stereo audio on one Radeon AI PRO R9700 in 5m 33s, with 16.7% lower latency than matched stock.","2026-09-09T07:30:00Z","paiton-minimax-h3-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-minimax-h3\u002F00-featured-minimax-h3-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-minimax-h3-radeon-ai-pro-r9700",[1384,1711,1712,1791,1792,1793],"Video Generation","MiniMax H3","ComfyUI",{"path":1795,"title":1796,"description":1797,"date":1798,"slug":1799,"image":1800,"originalUrl":1795,"categories":1801},"\u002Fblog\u002Fpaiton-flux2-klein-radeon-ai-pro-r9700","Local FLUX.2 klein on Radeon AI PRO R9700: Faster Image Generation with Less VRAM","Paiton generates 1024×1024 FLUX.2 klein images in 1.054 seconds on an R9700: 16.2% lower warm latency and 33.4% lower peak Torch allocation.","2026-09-07T09:00:00","paiton-flux2-klein-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-flux2-klein\u002Ffox-paiton.webp",[1384,1711,1712,1758,1802,1793],"FLUX",{"path":1804,"title":1805,"description":1806,"date":1807,"slug":1808,"image":1809,"originalUrl":1810,"categories":1811},"\u002Fblog\u002Fpaiton-ornith15-radeon-ai-pro-r9700","Ornith 1.5 at 44.6 tok\u002Fs on One Radeon AI PRO R9700","Paiton serves Ornith 1.5 35B A3B at 44.63 output tok\u002Fs on one Radeon AI PRO R9700, 27% faster than tuned stock vLLM, with 21.3% lower modeled cost.","2026-09-05T09:00:00","paiton-ornith15-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-ornith15\u002F00-featured-ornith15-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-ornith15-radeon-ai-pro-r9700",[1384,1812,1711,1813,1814,1815,1770,1816,1768,1817],"Artificial Intelligence","AI Inference","GPU Performance","Inference Latency","Large Language Models","Cost Efficiency",{"path":1819,"title":1820,"description":1821,"date":1822,"slug":1823,"image":1824,"originalUrl":1825,"categories":1826},"\u002Fblog\u002Fpaiton-qwen38-radeon-ai-pro-r9700","Paiton: 21% More Qwen3.8 Throughput on Radeon AI PRO R9700","Paiton serves AMD’s Qwen3.8 27B at 39.77 output tokens\u002Fs on one Radeon AI PRO R9700, delivering 21% more throughput and 17.4% lower modeled cost.","2026-09-04T09:00:00","paiton-qwen38-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-r9700\u002F00-featured-paiton-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-radeon-ai-pro-r9700",[1384,1812,1711,1813,1814,1815,1770,1816,1768,1817],{"path":1828,"title":1829,"description":1830,"date":1831,"slug":1832,"image":1833,"originalUrl":1727,"categories":1834},"\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt","AI Data Center Power Requirements: The GPU\u002FMW Illusion","Why do AI data-center proposals quote different GPU capacities? See how PUE, peak loads, storage, networking and cooling determine deployable compute.","2026-07-27T23:52:00","ai-data-center-power-requirements-gpu-per-megawatt","\u002Fasset\u002Fimages\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt\u002Fgpu-per-megawatt-illusion.webp",[1835,1836,1837,1838,1839,1840],"All","AI Infrastructure","Data Centers","ModFlex","HPC","AMD Helios",{"path":1842,"title":1843,"description":1844,"date":1845,"slug":1846,"image":1847,"originalUrl":1848,"categories":1849},"\u002Fblog\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x","Wan2.2 Video Generation: Paiton on AMD MI355X","Compare Wan2.2-T2V-A14B video generation on AMD MI355X with Paiton and NVIDIA B200 using Diffusers, and explore our diffusion optimization approach.","2026-06-10T14:04:04","paiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonwan2.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x\u002F",[1835,1812,1384,1850,1851,1852,1853,1854,1855,1856,1857,1858,1859,1860,1861,1862,1863,1864,1865,1866,1384,1867,1868,1869,1870,1871,1872],"14B","AMD","B200","Benchmarks","Blackwell","Compute","Diffusion","Eliovp","Generative AI","GPU","Hardware","Inference","Instinct","MI355x","NVidia","On-Premise","Optimization","Sovereign AI","T2V","Text-to-Video","Tuning","Video-Generation","Wan2.2",{"path":1874,"title":1875,"description":1876,"date":1877,"slug":1878,"image":1879,"originalUrl":1880,"categories":1881},"\u002Fblog\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure","ElioVP in De Tijd: Chip Optimization and Data Centers","Read about De Tijd's coverage of ElioVP, from its origins in chip optimization to its work on modular data centers and high-density cooling.","2026-02-10T20:48:12","from-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fphysicalnewspaper.webp","https:\u002F\u002Feliovp.com\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure\u002F",[1835,1812,1882,1883,1851,1884,1882,1864],"Modular DC","Uncategorized","De Tijd",{"path":1886,"title":1887,"description":1888,"date":1889,"slug":1890,"image":1891,"originalUrl":1892,"categories":1893},"\u002Fblog\u002Fprivacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit","AI Privacy: A Strategic Priority for Benelux Businesses","Explore generative AI privacy risks, trust, data retention and governance, and why Benelux businesses need a strategic approach to secure AI.","2026-01-29T13:51:11","privacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fheaderimage.webp","https:\u002F\u002Feliovp.com\u002Fprivacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit\u002F",[1835,1812,1894,1883,1895,1896,1897,1898,1899,1900,1901,1902,1857,1903,1858,1904,1905,1906],"Trending","AI Act","Anthropomorphism","AVG","Benelux","ChatGPT","Cybersecurity","Data Governance","Data Security","GDPR","Microsoft Copilot","Privacy","Shadow AI",{"path":1908,"title":1909,"description":1910,"date":1911,"slug":1912,"image":1913,"originalUrl":1914,"categories":1915},"\u002Fblog\u002Fitsme-bij-ons-is-het-its-not-me-en-dit-is-waarom","Why We Do Not Use itsme: Privacy and Data Sovereignty","Why ElioVP does not use itsme: our assessment of identity metadata, cloud dependence, data sovereignty and authentication risks.","2025-11-27T09:32:14","itsme-bij-ons-is-het-its-not-me-en-dit-is-waarom","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffrontimage.webp","https:\u002F\u002Feliovp.com\u002Fitsme-bij-ons-is-het-its-not-me-en-dit-is-waarom\u002F",[1835,1916,1917,1918,1900,1919,1920,1921,1903,1922,1923,1924,1905],"AWS","Belgian Mobile ID","Cloud Act","Data Sovereignty","Digital Identity","eIDAS","itsme","Liberty Global","MyGov.be",{"path":1926,"title":1927,"description":1928,"date":1929,"slug":1930,"image":1931,"originalUrl":1932,"categories":1933},"\u002Fblog\u002Ffield-report-the-reality-of-building-agentic-ai-in-2025","Field Report. The Reality of Building Agentic AI in 2025","Lessons from building on-premise AI agents in 2025 cover workflow design, observability, model training, hallucinations and GPU memory limits.","2025-11-25T14:03:39","field-report-the-reality-of-building-agentic-ai-in-2025","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffieldreport.webp","https:\u002F\u002Feliovp.com\u002Ffield-report-the-reality-of-building-agentic-ai-in-2025\u002F",[1835,1812,1934,1894,1935,1936,1937,1938,1939,1940,1941,1942,1867,1943],"Solutions","Agentic AI","AI Engineering","AI Strategy","Autonomous Agents","Enterprise AI","Local LLM","Model Fine-Tuning","On-Premise AI","VRAM Optimization",{"path":1945,"title":1946,"description":1947,"date":1948,"slug":1949,"image":1950,"originalUrl":1951,"categories":1952},"\u002Fblog\u002Fthe-synthetic-unicorn-bubble","The Synthetic Unicorn Bubble","An analysis of AI neocloud investment risks, examining circular financing, infrastructure claims, contract terms and due diligence.","2025-11-24T19:22:28","the-synthetic-unicorn-bubble","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fsyntheticunicorn.webp","https:\u002F\u002Feliovp.com\u002Fthe-synthetic-unicorn-bubble\u002F",[1835,1812,1894,1836,1953,1954,1955,1956,1957,1958,1959,1960,1961],"AI Neocloud","Circular Financing","GPU Cloud","Investment Risks","Startup Valuation","Synthetic Bubble","Tech Analysis","Vaporware","Venture Capital",{"path":1963,"title":1964,"description":1965,"date":1966,"slug":1967,"image":1968,"originalUrl":1969,"categories":1970},"\u002Fblog\u002Fbuilding-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power","NVIDIA GB300 NVL72: A Four-Month Modular Data Center Plan","Explore a modular data center design for NVIDIA GB300 NVL72, covering redundant power, hybrid cooling and a four-month deployment plan.","2025-11-20T14:10:19","building-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fsuperpodmodflexfrontimage.webp","https:\u002F\u002Feliovp.com\u002Fbuilding-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power\u002F",[1835,1882,1883,1971,1836,1972,1973,1974,1975,1976,1977,1978,1979],"150kW Rack","DLC","High Density","Liquid Cooling","Modular Data Center","NVIDIA Blackwell Ultra","NVIDIA GB300","NVL72","Rapid Deployment",{"path":1981,"title":1982,"description":1983,"date":1984,"slug":1985,"image":1986,"originalUrl":1987,"categories":1988},"\u002Fblog\u002Fwhy-cuda-translation-wont-unlock-amds-real-potential","Why “CUDA” Translation Won’t Unlock AMD’s Real Potential","Why CUDA compatibility is not the same as AMD performance: explore ROCm, HIP, kernel tuning and the case for hardware-specific optimization.","2025-11-12T14:48:37","why-cuda-translation-wont-unlock-amds-real-potential","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fchatgpt-image-nov-11-2025-09_16_10-pm-1.webp","https:\u002F\u002Feliovp.com\u002Fwhy-cuda-translation-wont-unlock-amds-real-potential\u002F",[1835,1812,1384,1883,1989,1812,1990,1991,1992,1993,1994,1995,1384,1996],"AMD MI300X","CUDA Translation","FP8","GPU Optimization","High Performance Computing","HIP","Kernel Tuning","ROCm",{"path":1998,"title":1999,"description":2000,"date":2001,"slug":2002,"image":2003,"originalUrl":2004,"categories":2005},"\u002Fblog\u002Fpaiton-the-simplest-way-to-supercharge-ai-inference","Paiton: The Simplest Way to Supercharge AI Inference","Learn how Paiton integrates with existing inference stacks, with AMD MI300X benchmark results and performance-per-dollar comparisons.","2025-11-11T10:31:22","paiton-the-simplest-way-to-supercharge-ai-inference","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaiton-powaaah.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-the-simplest-way-to-supercharge-ai-inference\u002F",[1835,1812,1384,1813,2006,1989,1817,2007,1770,1995,1384,2008,1768],"AMD Instinct","High Throughput","SGLang",{"path":2010,"title":2011,"description":2012,"date":2013,"slug":2014,"image":2015,"originalUrl":2016,"categories":2017},"\u002Fblog\u002Fstop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens","Paiton MoE Benchmarks: MI300X vs H200 and B200","Compare Qwen3-30B-A3B MoE inference with Paiton on MI300X against H200 and B200, including throughput and cost per million tokens.","2025-09-26T13:36:18","stop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fhulkvshulkpaitonwins.webp","https:\u002F\u002Feliovp.com\u002Fstop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens\u002F",[1835,1812,1384,2018,1989,2019,1770,2020,2021,2022,2023,1384,2024],"AI Benchmarks","Cost per Token","Mixture of Experts","MoE","NVIDIA B200","NVIDIA H200","Qwen3",{"path":2026,"title":2027,"description":2028,"date":2029,"slug":2030,"image":2031,"originalUrl":2032,"categories":2033},"\u002Fblog\u002Fagentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en","Local Agentic AI: From Inbox to Action","Local-first AI agents turn email, documents and images into tickets, reports and actions, using models tailored to your data and systems.","2025-09-16T13:09:00","agentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffrontfotoblog.webp","https:\u002F\u002Feliovp.com\u002Fagentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en\u002F",[1835,1812,1934,1883,1935,2034,2035,2036,2037,1940,1942,1867,2038,2039],"Damage Detection","Document Processing","Email Automation","Invoice Extraction","Ticket Automation","Workflow Automation",{"path":2041,"title":2042,"description":2043,"date":2044,"slug":2045,"image":2046,"originalUrl":2047,"categories":2048},"\u002Fblog\u002Fmi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach","MI300X FP8 Benchmarks: GPU Partitioning with Paiton","Explore Llama 3.1 8B FP8 benchmarks on partitioned MI300X GPUs with Paiton, comparing throughput and latency against NVIDIA H200 and B200.","2025-07-31T13:32:57","mi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fimage-2-1.webp","https:\u002F\u002Feliovp.com\u002Fmi300x-fp8-data%e2%80%91parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach\u002F",[1835,1812,1384,1883,2049,1851,1852,2050,2051,1863,1864,1384,1768],"AI","H200","MI300X",{"path":2053,"title":2054,"description":2055,"date":2056,"slug":2057,"image":2058,"originalUrl":2059,"categories":2060},"\u002Fblog\u002Fapplicable-ai-for-businesses","Applicable AI for Businesses","Explore ElioVP's approach to local AI for business workflows, including custom model training and automated damage detection for logistics.","2025-07-09T21:35:30","applicable-ai-for-businesses","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fscherm_afbeelding-2025-07-09-om-23.30.17.webp","https:\u002F\u002Feliovp.com\u002Fapplicable-ai-for-businesses\u002F",[1835,1812,1934,1935,2034,2035,2036,2037,1940,1942,1867,2038,2039],{"path":2062,"title":2063,"description":2064,"date":2065,"slug":2066,"image":31,"originalUrl":2067,"categories":2068},"\u002Fblog\u002Fintroducing-paitons-free-evaluation-models","Introducing Paiton’s Free Evaluation Models","Test Paiton with free evaluation models for AMD GPUs. Compare text, vision and image generation performance using your own workloads.","2025-07-07T11:26:13","introducing-paitons-free-evaluation-models","https:\u002F\u002Feliovp.com\u002Fintroducing-paitons-free-evaluation-models\u002F",[1835,1812,1384],{"path":2070,"title":2071,"description":2072,"date":2073,"slug":2074,"image":2075,"originalUrl":2076,"categories":2077},"\u002Fblog\u002Fpaiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b","Llama 3.1 405B: Faster Startup with Paiton on MI300X","See Paiton benchmarks for Llama 3.1 405B on eight AMD MI300X GPUs, covering model startup, tensor parallelism, throughput and latency.","2025-06-12T20:15:23","paiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fservingscreenshot.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b\u002F",[1835,1812,1384,1883,1813,1989,2078,1991,2079,2080,2081,1384,2082,2083],"Cold Start","Graph Compilation","Llama 3.1 405B","LLM Optimization","Startup Latency","Tensor Parallelism",{"path":2085,"title":2086,"description":2087,"date":2088,"slug":2089,"image":2090,"originalUrl":2091,"categories":2092},"\u002Fblog\u002Fpaiton-fp8-beats-nvidias-h200-on-amds-mi300x","Paiton FP8 Beats NVIDIA’s H200 on AMD’s MI300X","Compare Paiton on AMD MI300X with NVIDIA H200 for Llama 3.1 70B FP8, including throughput, first-token delay and latency across batch sizes.","2025-06-08T19:12:40","paiton-fp8-beats-nvidias-h200-on-amds-mi300x","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fblognewfp8.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-fp8-beats-nvidias-h200-on-amds-mi300x\u002F",[1835,1812,1384,1883,1989,2093,1939,1858,1814,1815,1816,2080,2094,2095],"Cold Start Optimization","Model Serving","vLLM Optimization",{"path":2097,"title":2098,"description":2099,"date":2100,"slug":2101,"image":2102,"originalUrl":2103,"categories":2104},"\u002Fblog\u002Fmi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm","MI300X vs H200 vs RX 7900 XTX vs Tenstorrent n300s with vLLM","Compare MI300X, H200, RX 7900 XTX and Tenstorrent n300s on Llama 3 8B with vLLM, including throughput, modeled token costs and hardware limits.","2025-05-09T14:03:58","mi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fcomparisontenstor.webp","https:\u002F\u002Feliovp.com\u002Fmi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm\u002F",[1835,1812,1384,1934,1883,1851,2051,1864,2105,2106],"RX7900XTX","tenstorrent",{"path":2108,"title":2109,"description":2110,"date":2111,"slug":2112,"image":2113,"originalUrl":2114,"categories":2115},"\u002Fblog\u002Fclusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights","ClusterP&L: Financial Modeling for GPU Clusters","Explore how ClusterP&L models GPU cluster costs, profitability and investment scenarios, with ROI metrics, risk simulations and exportable reports.","2025-05-03T10:52:22","clusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fcomparisonscenarios.webp","https:\u002F\u002Feliovp.com\u002Fclusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights\u002F",[1835,1812,1882,1934,1852,2050,2116,1864,2117],"MI325x","pnl calculator",{"path":2119,"title":2120,"description":2121,"date":2122,"slug":2123,"image":2124,"originalUrl":2125,"categories":2126},"\u002Fblog\u002Fcranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200","AMD MI300X vs. NVIDIA H200: Qwen3-32B with Paiton","Compare Qwen3-32B benchmarks on Paiton-optimized AMD MI300X and NVIDIA H200, covering throughput, latency and hardware costs.","2025-05-02T21:10:30","cranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002F3ac59a73-2466-4422-b7e5-ef2e4a8ca58e.webp","https:\u002F\u002Feliovp.com\u002Fcranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200\u002F",[1835,1812,1384,2049,1851,2050,2127,1864,1384,1768],"MI300",{"path":2129,"title":2130,"description":2131,"date":2132,"slug":2133,"image":2134,"originalUrl":2135,"categories":2136},"\u002Fblog\u002Fpower-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw","Modular Data Centers for NVIDIA NVL: 1 to 2 MW","Explore modular data center designs for NVIDIA NVL systems, covering power capacity, liquid cooling, redundancy and deployment planning.","2025-05-02T14:09:59","power-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Feliovp_critical-1mw-pod_rev-2_transparent.webp","https:\u002F\u002Feliovp.com\u002Fpower-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw\u002F",[1835,1882,2137,1836,1973,1839,1974,1975,2138,2139,1978,2140],"1-2MW Data Center","NVIDIA Blackwell","NVIDIA NVL","Precision Cooling",{"path":2142,"title":2143,"description":2144,"date":2145,"slug":2146,"image":2147,"originalUrl":2148,"categories":2149},"\u002Fblog\u002Fexamining-ai-agents-in-the-medical-field-ai-that-speaks-dicom","Examining AI agents in the medical field: AI that speaks DICOM","Explore a local AI agent demo for DICOM workflows, from patient and study retrieval to a comparison of vision models using anonymized medical images.","2025-04-11T14:45:48","examining-ai-agents-in-the-medical-field-ai-that-speaks-dicom","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fhealthcareblog-1.webp","https:\u002F\u002Feliovp.com\u002Fexamining-ai-agents-in-the-medical-field-ai-that-speaks-dicom\u002F",[1835,1812,1934,1883,2049,1851,2150],"Healthcare",{"path":2152,"title":2153,"description":2154,"date":2155,"slug":2156,"image":2157,"originalUrl":2158,"categories":2159},"\u002Fblog\u002Feliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs","U.S. Tariffs and AI Supply Chain Resilience: April 2025","Read ElioVP's April 2025 perspective on U.S. tariffs and supply chain resilience for AI servers, HPC systems and modular data centers.","2025-04-04T10:01:27","eliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ftariffsshipping.webp","https:\u002F\u002Feliovp.com\u002Feliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs\u002F",[1835,1894,2049,1851,2160,2161,2162,2163],"import","Taiwan","Tariffs","Trump",{"path":2165,"title":2166,"description":2167,"date":2168,"slug":2169,"image":2170,"originalUrl":2171,"categories":2172},"\u002Fblog\u002Fwhy-ai-agents-are-the-future","Why AI Agents Are the Future","Explore AI agents for ERP, CRM, finance and customer support, with practical use cases and a path from workflow assessment to pilot and deployment.","2025-03-23T22:06:59","why-ai-agents-are-the-future","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ferp2.jpeg","https:\u002F\u002Feliovp.com\u002Fwhy-ai-agents-are-the-future\u002F",[1835,1812,1934,2049,2173,2174],"AI Agents","ERP",{"path":2176,"title":2177,"description":2178,"date":2179,"slug":2180,"image":2181,"originalUrl":2182,"categories":2183},"\u002Fblog\u002Fthe-rise-of-open-source-ai-model-optimization","The Rise of Open-Source AI Model Optimization","Explore open-source AI optimization trends, from quantization and mixture-of-experts models to hardware-aware tuning, RAG and edge deployment.","2025-03-22T20:59:23","the-rise-of-open-source-ai-model-optimization","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Friseofopensource.jpeg","https:\u002F\u002Feliovp.com\u002Fthe-rise-of-open-source-ai-model-optimization\u002F",[1835,1812,1894,2184,1851,1859,1864],"AI news",{"path":2186,"title":2187,"description":2188,"date":2189,"slug":2190,"image":2191,"originalUrl":2192,"categories":2193},"\u002Fblog\u002Fintroducing-our-benchmarking-tool-powered-by-dstack","Introducing Our Benchmarking Tool: Powered by dstack","Explore our dstack-powered tool for reproducible vLLM benchmarks, automated parameter sweeps and performance reports across local and cloud GPUs.","2025-03-20T14:21:59","introducing-our-benchmarking-tool-powered-by-dstack","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fbenchmarktool.jpeg","https:\u002F\u002Feliovp.com\u002Fintroducing-our-benchmarking-tool-powered-by-dstack\u002F",[1835,1812,1384,2049,1851,2194,2195,2051,1384],"benchmark","LLM",{"path":2197,"title":2198,"description":2199,"date":2200,"slug":2201,"image":2202,"originalUrl":2203,"categories":2204},"\u002Fblog\u002Foptimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200","Optimizing QwQ-32B (by Qwen): AMD MI300X vs. NVIDIA H200","Compare QwQ-32B throughput and latency on AMD MI300X with Paiton and NVIDIA H200, from small batches to higher concurrency.","2025-03-19T21:41:44","optimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaiton4.jpeg","https:\u002F\u002Feliovp.com\u002Foptimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200\u002F",[1835,1812,1384],{"path":2206,"title":2207,"description":2208,"date":2209,"slug":2210,"image":2211,"originalUrl":2212,"categories":2213},"\u002Fblog\u002Feliovp-featured-on-amd-tech-talk-podcast","Eliovp Featured on AMD “Tech Talk” Podcast","Listen to Elio Van Puyvelde and Jim Greene on AMD's Tech Talk podcast, discussing ElioVP's origins and its AI hardware and software services.","2025-03-19T07:53:39","eliovp-featured-on-amd-tech-talk-podcast","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ftechtalkjimgreene.jpeg","https:\u002F\u002Feliovp.com\u002Feliovp-featured-on-amd-tech-talk-podcast\u002F",[1835,1851,2214,2215,2216],"Jim Greene","Podcast","Tech Talk",{"path":2218,"title":2219,"description":2220,"date":2221,"slug":2222,"image":2223,"originalUrl":2224,"categories":2225},"\u002Fblog\u002Ffurther-optimizing-amd-powered-inference-with-paiton","Further Optimizing AMD-Powered Inference with Paiton","Explore Paiton's DeepSeek R1 Distill Llama 8B benchmarks on AMD MI300X, focusing on throughput and first-token latency at smaller batch sizes.","2025-03-13T06:18:30","further-optimizing-amd-powered-inference-with-paiton","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost3.webp","https:\u002F\u002Feliovp.com\u002Ffurther-optimizing-amd-powered-inference-with-paiton\u002F",[1835,1812,1384,1851,2226,2227,2050,2051,2116,1384,1768],"Deepseek","H100",{"path":2229,"title":2230,"description":2231,"date":2232,"slug":2233,"image":2234,"originalUrl":2235,"categories":2236},"\u002Fblog\u002Fa-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b","Paiton Benchmarks: DeepSeek R1 Distill Llama 3.1 8B","Compare stock and Paiton-optimized DeepSeek R1 Distill Llama 3.1 8B on AMD MI300X, with throughput and latency benchmarks across batch sizes.","2025-01-31T09:11:02","a-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost2.webp","https:\u002F\u002Feliovp.com\u002Fa-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b\u002F",[1835,1812,1384,1851,2226,2227,2050,2051,2116,1384,1768],{"path":2238,"title":2239,"description":2240,"date":2241,"slug":2242,"image":2243,"originalUrl":2244,"categories":2245},"\u002Fblog\u002Fai-model-optimization-with-paiton","AI Model Optimization with Paiton","Learn how Paiton uses model compilation, custom kernels and kernel fusion to optimize AI inference on AMD GPUs.","2025-01-30T19:53:25","ai-model-optimization-with-paiton","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost1.webp","https:\u002F\u002Feliovp.com\u002Fai-model-optimization-with-paiton\u002F",[1835,1812,1384,1851,2227,2050,2051,2116,1384,1768],1791835012262]