[{"data":1,"prerenderedAt":1836},["ShallowReactive",2],{"blog-post-en-\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp":3,"blog-posts-sidebar-en":1301},{"id":4,"title":5,"body":6,"categories":1282,"date":1287,"description":1288,"extension":1289,"heading":1290,"image":1291,"meta":1292,"navigation":776,"originalUrl":1293,"path":1294,"seo":1295,"slug":1296,"socialImage":1297,"stem":1298,"updated":1299,"video":1299,"__hash__":1300},"blog\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp.md","Qwen3.8 Flash Next: 3-bit near 16-bit accuracy, with MTP prefix caching",{"type":7,"value":8,"toc":1269},"minimark",[9,16,48,63,68,87,91,96,193,214,231,236,283,302,306,323,327,336,339,344,427,440,460,483,498,502,511,514,519,590,616,620,638,654,658,673,690,694,715,726,872,900,910,929,944,948,971,984,1265],[10,11,12],"p",{},[13,14,15],"em",{},"Illustration: AI-generated editorial artwork, not a hardware photograph or benchmark capture.",[10,17,18,19,23,24,27,28,40],{},"Better answers, less waiting between turns. Our 24 October Paiton release brings ",[20,21,22],"strong",{},"new 3-bit Qwen3.8 Flash Next weights close to the 16-bit reference on three task benchmarks",", plus native prefix caching with multi-token prediction (MTP) for long conversations on ",[20,25,26],{},"two Radeon AI PRO R9700 cards",". A separate accuracy update also improves our 27B model on one card.",[29,30,31],"sup",{},[32,33,39],"a",{"href":34,"ariaDescribedBy":35,"dataFootnoteRef":37,"id":38},"#user-content-fn-threebit",[36],"footnote-label","","user-content-fnref-threebit","1",[29,41,42],{},[32,43,47],{"href":44,"ariaDescribedBy":45,"dataFootnoteRef":37,"id":46},"#user-content-fn-twentyseven",[36],"user-content-fnref-twentyseven","2",[10,49,50,51,54,55],{},"The latest 3-bit cache mode resumes a repeated ",[20,52,53],{},"200K-token prompt in 0.68 seconds",", versus 25.81 seconds cold, while keeping speculative decoding on. These are cache-hit measurements, not the wait for a new document.",[29,56,57],{},[32,58,62],{"href":59,"ariaDescribedBy":60,"dataFootnoteRef":37,"id":61},"#user-content-fn-bench",[36],"user-content-fnref-bench","3",[64,65,67],"h2",{"id":66},"accuracy-close-task-scores-at-3-bit","Accuracy: close task scores at 3-bit",[10,69,70,71,75,76,80,81],{},"The new weights improve the calibration behind our ",[32,72,74],{"href":73},"\u002Fblog\u002Fpaiton-qwen38-flash-next-radeon-ai-pro-r9700","original 3-bit Flash Next release",", without changing its format, size or kernels. Here they are alongside the ",[32,77,79],{"href":78},"\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700","4-bit release"," and original 16-bit model.",[29,82,83],{},[32,84,39],{"href":34,"ariaDescribedBy":85,"dataFootnoteRef":37,"id":86},[36],"user-content-fnref-threebit-2",[88,89],"qwen-accuracy-mtp-figure",{"kind":90},"accuracy",[10,92,93],{},[13,94,95],{},"Accuracy; higher is better. Text tasks ran through the released server configurations, not the native throughput setup below. AI2D emulates served arithmetic and is shown separately. The 3-bit results are from 24 October; 4-bit and 16-bit references are from 12 October.",[97,98,99,122],"table",{},[100,101,102],"thead",{},[103,104,105,109,113,116,119],"tr",{},[106,107,108],"th",{},"Task",[106,110,112],{"align":111},"right","Questions",[106,114,115],{"align":111},"16-bit",[106,117,118],{"align":111},"New 3-bit",[106,120,121],{"align":111},"4-bit",[123,124,125,143,159,176],"tbody",{},[103,126,127,131,134,137,140],{},[128,129,130],"td",{},"GSM8K, math",[128,132,133],{"align":111},"1,319",[128,135,136],{"align":111},"96.9%",[128,138,139],{"align":111},"97.4%",[128,141,142],{"align":111},"96.8%",[103,144,145,148,151,154,157],{},[128,146,147],{},"HumanEval, Python pass@1",[128,149,150],{"align":111},"164",[128,152,153],{"align":111},"96.3%",[128,155,156],{"align":111},"94.5%",[128,158,153],{"align":111},[103,160,161,164,167,170,173],{},[128,162,163],{},"MMLU-Pro, knowledge",[128,165,166],{"align":111},"1,400",[128,168,169],{"align":111},"84.7%",[128,171,172],{"align":111},"83.6%",[128,174,175],{"align":111},"83.7%",[103,177,178,181,184,187,190],{},[128,179,180],{},"AI2D, science diagrams",[128,182,183],{"align":111},"600",[128,185,186],{"align":111},"85.3%",[128,188,189],{"align":111},"81.8%",[128,191,192],{"align":111},"Not reported",[10,194,195,196,199,200,206],{},"None of the three text-task differences from 16-bit is statistically significant in the paired tests. That ",[20,197,198],{},"does not establish equivalence",": HumanEval is 1.8 percentage points lower, with a paired 95% interval of −5.0 to +1.3 points. The 16-bit reference itself moved by up to 1.2 points between two seed sets.",[29,201,202],{},[32,203,39],{"href":34,"ariaDescribedBy":204,"dataFootnoteRef":37,"id":205},[36],"user-content-fnref-threebit-3",[29,207,208],{},[32,209,213],{"href":210,"ariaDescribedBy":211,"dataFootnoteRef":37,"id":212},"#user-content-fn-fourbit",[36],"user-content-fnref-fourbit","4",[10,215,216,217,223],{},"Vision remains behind the reference. The AI2D evaluation does not establish native-worker image-input support: the released 3-bit modes are text-only.",[29,218,219],{},[32,220,39],{"href":34,"ariaDescribedBy":221,"dataFootnoteRef":37,"id":222},[36],"user-content-fnref-threebit-4",[29,224,225],{},[32,226,230],{"href":227,"ariaDescribedBy":228,"dataFootnoteRef":37,"id":229},"#user-content-fn-release",[36],"user-content-fnref-release","5",[232,233,235],"h3",{"id":234},"how-close-are-the-token-predictions","How close are the token predictions?",[97,237,238,249],{},[100,239,240],{},[103,241,242,245,247],{},[106,243,244],{},"Held-out check against 16-bit",[106,246,118],{"align":111},[106,248,121],{"align":111},[123,250,251,262,273],{},[103,252,253,256,259],{},[128,254,255],{},"Full-vocabulary KL divergence, lower is better",[128,257,258],{"align":111},"0.055",[128,260,261],{"align":111},"0.031",[103,263,264,267,270],{},[128,265,266],{},"Relative long-document negative log-likelihood increase",[128,268,269],{"align":111},"+0.35%",[128,271,272],{"align":111},"+0.16%",[103,274,275,278,281],{},[128,276,277],{},"Needle retrieval at 131,072 tokens",[128,279,280],{"align":111},"40 \u002F 40",[128,282,280],{"align":111},[10,284,285,286,289,290,296],{},"KL measures differences in next-token probability distributions; zero means identical distributions. Negative log-likelihood measures how well predictions fit observed text. ",[20,287,288],{},"These percentages are relative NLL changes, not perplexity changes."," Distribution checks emulate served arithmetic; the separate retrieval checks use the served model. Neither replaces broad task evaluation.",[29,291,292],{},[32,293,39],{"href":34,"ariaDescribedBy":294,"dataFootnoteRef":37,"id":295},[36],"user-content-fnref-threebit-5",[29,297,298],{},[32,299,213],{"href":210,"ariaDescribedBy":300,"dataFootnoteRef":37,"id":301},[36],"user-content-fnref-fourbit-2",[232,303,305],{"id":304},"what-changed-in-the-weights","What changed in the weights",[10,307,308,309,312,313,316,317],{},"The recipe combines separate calibration statistics for each expert's output projection, a more varied calibration set of about ",[20,310,311],{},"four million tokens",", and retained ",[20,314,315],{},"8-bit shared layers"," around the 3-bit experts. Only the rounding changes; the packed format and runtime kernels stay the same.",[29,318,319],{},[32,320,39],{"href":34,"ariaDescribedBy":321,"dataFootnoteRef":37,"id":322},[36],"user-content-fnref-threebit-6",[64,324,326],{"id":325},"prefix-caching-with-mtp-still-on","Prefix caching with MTP still on",[10,328,329,330],{},"Coding agents resend their conversation on every turn. Prefix caching reuses the unchanged part; MTP drafts several next tokens for the target model to verify. The native worker now combines both. It runs the model step without PyTorch on the GPU, while vLLM still provides the API and scheduling.",[29,331,332],{},[32,333,230],{"href":227,"ariaDescribedBy":334,"dataFootnoteRef":37,"id":335},[36],"user-content-fnref-release-2",[88,337],{"kind":338},"prefix-cache",[10,340,341],{},[13,342,343],{},"Time to first token in separate repeated-prompt checks; lower is better. A hit needs an unchanged prefix still available in GPU or RAM cache. These are not cold-cache prefill-throughput benchmarks.",[97,345,346,359],{},[100,347,348],{},[103,349,350,353,356],{},[106,351,352],{},"Configuration and repeated prompt",[106,354,355],{"align":111},"Cold",[106,357,358],{"align":111},"Cache hit",[123,360,361,372,383,394,405,416],{},[103,362,363,366,369],{},[128,364,365],{},"4-bit native MTP cache, 32K",[128,367,368],{"align":111},"5.25 s",[128,370,371],{"align":111},"0.69 s",[103,373,374,377,380],{},[128,375,376],{},"4-bit native MTP cache, 90K",[128,378,379],{"align":111},"14.68 s",[128,381,382],{"align":111},"0.78 s",[103,384,385,388,391],{},[128,386,387],{},"4-bit native MTP cache, 200K",[128,389,390],{"align":111},"33.62 s",[128,392,393],{"align":111},"0.85 s",[103,395,396,399,402],{},[128,397,398],{},"4-bit idle session restored from RAM, 145K–200K",[128,400,401],{"align":111},"21–30 s",[128,403,404],{"align":111},"0.98–1.17 s",[103,406,407,410,413],{},[128,408,409],{},"New 3-bit native MTP cache, 90K",[128,411,412],{"align":111},"11.09 s",[128,414,415],{"align":111},"0.60 s",[103,417,418,421,424],{},[128,419,420],{},"New 3-bit native MTP cache, 200K",[128,422,423],{"align":111},"25.81 s",[128,425,426],{"align":111},"0.68 s",[10,428,429,430,433,434],{},"Tested requests served alone produce identical cached and cold output; concurrent batching can change tokens on either path. An optional RAM tier retains evicted sessions and sizes itself to available memory. The 4-bit cache configuration generates at ",[20,431,432],{},"206.9 tokens\u002Fs with that tier on",".",[29,435,436],{},[32,437,62],{"href":59,"ariaDescribedBy":438,"dataFootnoteRef":37,"id":439},[36],"user-content-fnref-bench-2",[10,441,442,443,449,450,453,454],{},"The newer ",[20,444,445],{},[446,447,448],"code",{},"3bit-long-cache-mtp-native"," mode extends this to a ",[20,451,452],{},"262,144-token window",": 213.5 tokens\u002Fs with the RAM tier on, 250.0 off. Its cache and decode figures are separate validation measurements, not reruns of the main release comparison below. Prompt, formatting and answer must fit the window together.",[29,455,456],{},[32,457,62],{"href":59,"ariaDescribedBy":458,"dataFootnoteRef":37,"id":459},[36],"user-content-fnref-bench-3",[10,461,462,463,468,469,472,473,476,477],{},"For smaller hosts, ",[20,464,465],{},[446,466,467],{},"--mode decode --fn-worker native --prefix-caching"," supports ",[20,470,471],{},"64 GB RAM",", a 98,304-token window and about 101 tokens\u002Fs, ",[20,474,475],{},"without MTP",". It is a different mode, not the configuration behind the 213.5 tokens\u002Fs result.",[29,478,479],{},[32,480,230],{"href":227,"ariaDescribedBy":481,"dataFootnoteRef":37,"id":482},[36],"user-content-fnref-release-3",[10,484,485,486,492],{},"One near-capacity test with 2.05 million idle-session tokens resumed cold despite expected RAM hits. Output stayed correct; the issue remains open. Also, the 3-bit model is more prone to early endings beyond roughly 28K context in tested sessions. The public guidance prefers 4-bit for long coding conversations.",[29,487,488],{},[32,489,62],{"href":59,"ariaDescribedBy":490,"dataFootnoteRef":37,"id":491},[36],"user-content-fnref-bench-4",[29,493,494],{},[32,495,39],{"href":34,"ariaDescribedBy":496,"dataFootnoteRef":37,"id":497},[36],"user-content-fnref-threebit-7",[64,499,501],{"id":500},"generation-speed-on-two-workstation-cards","Generation speed on two workstation cards",[10,503,504,505],{},"The accuracy update changes the weight values, not the amount of data read or the kernels. The published native-worker throughput remains in the same performance range.",[29,506,507],{},[32,508,62],{"href":59,"ariaDescribedBy":509,"dataFootnoteRef":37,"id":510},[36],"user-content-fnref-bench-5",[88,512],{"kind":513},"throughput",[10,515,516],{},[13,517,518],{},"Two R9700 cards, native worker, prefix caching off. Higher is better. Single-request weighted decode and aggregate concurrent throughput are different workloads, not a per-user scaling comparison.",[97,520,521,533],{},[100,522,523],{},[103,524,525,528,531],{},[106,526,527],{},"Measurement",[106,529,530],{"align":111},"3-bit",[106,532,121],{"align":111},[123,534,535,546,557,568,579],{},[103,536,537,540,543],{},[128,538,539],{},"BetterBench weighted decode, one request",[128,541,542],{"align":111},"258.7 tokens\u002Fs",[128,544,545],{"align":111},"245.3 tokens\u002Fs",[103,547,548,551,554],{},[128,549,550],{},"BetterBench, eight requests, aggregate",[128,552,553],{"align":111},"687.5 tokens\u002Fs",[128,555,556],{"align":111},"566.7 tokens\u002Fs",[103,558,559,562,565],{},[128,560,561],{},"Reading a nominal 64K prompt",[128,563,564],{"align":111},"8,433 input tokens\u002Fs",[128,566,567],{"align":111},"7,788 input tokens\u002Fs",[103,569,570,573,576],{},[128,571,572],{},"LocalMaxxing, one user",[128,574,575],{"align":111},"220.3 tokens\u002Fs",[128,577,578],{"align":111},"200.5 tokens\u002Fs",[103,580,581,584,587],{},[128,582,583],{},"LocalMaxxing, eight users, aggregate",[128,585,586],{"align":111},"781.3 tokens\u002Fs",[128,588,589],{"align":111},"680.2 tokens\u002Fs",[10,591,592,595,596,599,600,603,604,610],{},[20,593,594],{},"These are opt-in native-worker results."," Ordinary ",[446,597,598],{},"decode"," and ",[446,601,602],{},"4bit"," commands still select the torch worker unless requested otherwise. Neither the cache modes nor their RAM tiers are included in this table.",[29,605,606],{},[32,607,230],{"href":227,"ariaDescribedBy":608,"dataFootnoteRef":37,"id":609},[36],"user-content-fnref-release-4",[29,611,612],{},[32,613,62],{"href":59,"ariaDescribedBy":614,"dataFootnoteRef":37,"id":615},[36],"user-content-fnref-bench-6",[64,617,619],{"id":618},"a-second-accuracy-update-for-qwen38-27b","A second accuracy update for Qwen3.8 27B",[10,621,622,623,627,628,631,632],{},"Our ",[32,624,626],{"href":625},"\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700","3-bit 27B release"," runs on one R9700. This update keeps the 3-bit codes fixed and further trains their scale factors toward the 16-bit model, using ",[20,629,630],{},"4.2 million new tokens"," separate from earlier calibration and reported benchmarks.",[29,633,634],{},[32,635,47],{"href":44,"ariaDescribedBy":636,"dataFootnoteRef":37,"id":637},[36],"user-content-fnref-twentyseven-2",[10,639,640,641,644,645,433,648],{},"Served KL falls ",[20,642,643],{},"3.0%",", with a 95% interval of 0.4% to 6.4%. Long-document predictions improve across 8K–32K buckets; the 131K check finds no worsening, with gains from 64K onward. Format, size and kernel cost are unchanged; measured throughput can vary with speculative acceptance. New task checks show no significant change, but are ",[20,646,647],{},"weight-only evaluations, not fresh served-model task runs",[29,649,650],{},[32,651,47],{"href":44,"ariaDescribedBy":652,"dataFootnoteRef":37,"id":653},[36],"user-content-fnref-twentyseven-3",[64,655,657],{"id":656},"how-to-read-the-measurements","How to read the measurements",[10,659,660,661,667],{},"Flash Next text tasks use thinking on, temperature 1.0, top-p 0.95, top-k 20 and a 16,384-token answer limit. Prompts and per-question seeds match; capped answers count as wrong. GSM8K and HumanEval use their full test sets; MMLU-Pro uses 100 questions from each of 14 subjects. The original 16-bit weights ran twice on a larger AMD server. Comparisons use paired confidence intervals and exact McNemar tests.",[29,662,663],{},[32,664,39],{"href":34,"ariaDescribedBy":665,"dataFootnoteRef":37,"id":666},[36],"user-content-fnref-threebit-8",[29,668,669],{},[32,670,213],{"href":210,"ariaDescribedBy":671,"dataFootnoteRef":37,"id":672},[36],"user-content-fnref-fourbit-3",[10,674,675,676,679,680,683,684],{},"Performance uses ",[20,677,678],{},"BetterBench 0.6.0, standard profile",", one run per configuration, and ",[20,681,682],{},"LocalMaxxing v0.1.54",", retaining the median single-user-throughput run of three after warm-up. These are our measurements, not independent reproductions.",[29,685,686],{},[32,687,62],{"href":59,"ariaDescribedBy":688,"dataFootnoteRef":37,"id":689},[36],"user-content-fnref-bench-7",[64,691,693],{"id":692},"try-the-release","Try the release",[10,695,696,697,700,701,704,705,708,709],{},"Use ",[20,698,699],{},"Linux x86-64, two R9700 cards, a ROCm 10 host driver, Docker GPU access and Python 3.10+",". Fast 3-bit native and 3-bit MTP caching need roughly ",[20,702,703],{},"128 GB+ RAM and 60 GiB free GTT","; 4-bit native needs about a ",[20,706,707],{},"98 GiB GTT limit",". Preflight checks the actual host. Allow 116 GB for 3-bit weights or 133 GB for 4-bit, plus the roughly 56 GB native image.",[29,710,711],{},[32,712,230],{"href":227,"ariaDescribedBy":713,"dataFootnoteRef":37,"id":714},[36],"user-content-fnref-release-5",[10,716,717,718],{},"Use a fresh checkout if you have local plugin changes. The pinned launcher selects its runtime and model revisions, downloads the updated 3-bit weights into a new folder and leaves the old download intact.",[29,719,720],{},[32,721,725],{"href":722,"ariaDescribedBy":723,"dataFootnoteRef":37,"id":724},"#user-content-fn-launcher",[36],"user-content-fnref-launcher","6",[727,728,732],"pre",{"className":729,"code":730,"language":731,"meta":37,"style":37},"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 4882e36e455c1e657175db910367c992837c72a5\n\npython3 -m venv .venv\n. .venv\u002Fbin\u002Factivate\npython3 -m pip install huggingface_hub\n\npython3 models\u002FQwen3.8-Flash-Next\u002Flaunch-flashnext.py \\\n  --mode 3bit-long-cache-mtp-native --host 127.0.0.1 --dry-run\npython3 models\u002FQwen3.8-Flash-Next\u002Flaunch-flashnext.py \\\n  --mode 3bit-long-cache-mtp-native --host 127.0.0.1\n","bash",[446,733,734,750,760,771,778,793,801,817,822,833,851,860],{"__ignoreMap":37},[735,736,739,743,747],"span",{"class":737,"line":738},"line",1,[735,740,742],{"class":741},"sScJk","git",[735,744,746],{"class":745},"sZZnC"," clone",[735,748,749],{"class":745}," https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin.git\n",[735,751,753,757],{"class":737,"line":752},2,[735,754,756],{"class":755},"sj4cs","cd",[735,758,759],{"class":745}," paiton-vllm-plugin\n",[735,761,763,765,768],{"class":737,"line":762},3,[735,764,742],{"class":741},[735,766,767],{"class":745}," checkout",[735,769,770],{"class":745}," 4882e36e455c1e657175db910367c992837c72a5\n",[735,772,774],{"class":737,"line":773},4,[735,775,777],{"emptyLinePlaceholder":776},true,"\n",[735,779,781,784,787,790],{"class":737,"line":780},5,[735,782,783],{"class":741},"python3",[735,785,786],{"class":755}," -m",[735,788,789],{"class":745}," venv",[735,791,792],{"class":745}," .venv\n",[735,794,796,798],{"class":737,"line":795},6,[735,797,433],{"class":755},[735,799,800],{"class":745}," .venv\u002Fbin\u002Factivate\n",[735,802,804,806,808,811,814],{"class":737,"line":803},7,[735,805,783],{"class":741},[735,807,786],{"class":755},[735,809,810],{"class":745}," pip",[735,812,813],{"class":745}," install",[735,815,816],{"class":745}," huggingface_hub\n",[735,818,820],{"class":737,"line":819},8,[735,821,777],{"emptyLinePlaceholder":776},[735,823,825,827,830],{"class":737,"line":824},9,[735,826,783],{"class":741},[735,828,829],{"class":745}," models\u002FQwen3.8-Flash-Next\u002Flaunch-flashnext.py",[735,831,832],{"class":755}," \\\n",[735,834,836,839,842,845,848],{"class":737,"line":835},10,[735,837,838],{"class":755},"  --mode",[735,840,841],{"class":745}," 3bit-long-cache-mtp-native",[735,843,844],{"class":755}," --host",[735,846,847],{"class":755}," 127.0.0.1",[735,849,850],{"class":755}," --dry-run\n",[735,852,854,856,858],{"class":737,"line":853},11,[735,855,783],{"class":741},[735,857,829],{"class":745},[735,859,832],{"class":755},[735,861,863,865,867,869],{"class":737,"line":862},12,[735,864,838],{"class":755},[735,866,841],{"class":745},[735,868,844],{"class":755},[735,870,871],{"class":755}," 127.0.0.1\n",[10,873,874,875,878,879,882,883,886,887,890,891,433,894],{},"For 4-bit MTP caching, use ",[446,876,877],{},"bash models\u002FQwen3.8-Flash-Next\u002Frun-flashnext-4bit.sh"," with ",[446,880,881],{},"--fn-worker native --prefix-caching --host 127.0.0.1"," instead. Wait for ",[446,884,885],{},"launcher: READY","; the API is ",[446,888,889],{},"http:\u002F\u002F127.0.0.1:18982\u002Fv1",", model ",[446,892,893],{},"Qwen3.8-Flash-Next",[29,895,896],{},[32,897,230],{"href":227,"ariaDescribedBy":898,"dataFootnoteRef":37,"id":899},[36],"user-content-fnref-release-6",[10,901,902,903,909],{},"For 27B, follow the ",[32,904,908],{"href":905,"rel":906},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002F4882e36e455c1e657175db910367c992837c72a5\u002Fmodels\u002FQwen3.8-MXFP4-DFlash2\u002FREADME.md",[907],"nofollow","pinned setup and upgrade guide",". Download its new weight revision explicitly; updating the launcher alone does not update those weights.",[10,911,912,913,918,919,922,923],{},"The server's default binding exposes all interfaces without an API key. Keep ",[20,914,915],{},[446,916,917],{},"--host 127.0.0.1"," for local use. For remote clients, protect access with a firewall and ",[446,920,921],{},"PAITON_API_KEY","; avoid placing a key directly in command arguments. Stop the running server before selecting another mode.",[29,924,925],{},[32,926,230],{"href":227,"ariaDescribedBy":927,"dataFootnoteRef":37,"id":928},[36],"user-content-fnref-release-7",[10,930,931,932,937,938],{},"Native modes reject structured outputs, strict\u002Frequired tool calls and other API options; ordinary automatic tool calling works. Check the ",[32,933,936],{"href":934,"rel":935},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002F4882e36e455c1e657175db910367c992837c72a5\u002Fmodels\u002FQwen3.8-Flash-Next\u002FREADME.md#native-worker-opt-in",[907],"compatibility notes"," before connecting an agent.",[29,939,940],{},[32,941,230],{"href":227,"ariaDescribedBy":942,"dataFootnoteRef":37,"id":943},[36],"user-content-fnref-release-8",[64,945,947],{"id":946},"credits-and-licensing","Credits and licensing",[10,949,950,951,954,955,958,959,965],{},"Qwen supplies the models, vLLM the serving framework, and BetterBench and LocalMaxxing the benchmark tools. Flash Next weights use ",[20,952,953],{},"Qwen Community License 1.0",", with commercial-use conditions; the 27B weights use ",[20,956,957],{},"Apache 2.0",". Review the model licenses and calibration attributions before deployment.",[29,960,961],{},[32,962,39],{"href":34,"ariaDescribedBy":963,"dataFootnoteRef":37,"id":964},[36],"user-content-fnref-threebit-9",[29,966,967],{},[32,968,47],{"href":44,"ariaDescribedBy":969,"dataFootnoteRef":37,"id":970},[36],"user-content-fnref-twentyseven-4",[10,972,973,974,978,979,983],{},"Explore ",[32,975,977],{"href":976},"\u002Fproducts\u002Fpaiton","Paiton"," for the runtime, or ",[32,980,982],{"href":981},"\u002Fproducts\u002Fsovereign-ai","sovereign AI"," for a locally controlled deployment. For the underlying calibration, modes and reproducibility details, follow the pinned sources below.",[985,986,989,994],"section",{"className":987,"dataFootnotes":37},[988],"footnotes",[64,990,993],{"className":991,"id":36},[992],"sr-only","Footnotes",[995,996,997,1074,1108,1163,1190,1252],"ol",{},[998,999,1001,1006,1007,1014,1015,1014,1022,1014,1029,1014,1036,1014,1043,1014,1050,1014,1058,1014,1066],"li",{"id":1000},"user-content-fn-threebit",[32,1002,1005],{"href":1003,"rel":1004},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FQwen3.8-Flash-Next-W3A8-Paiton-RDNA4\u002Fblob\u002Fa100b3aee53b2cf2c9636650387936aab22dd5d7\u002FREADME.md",[907],"Flash Next W3A8 model card, task accuracy and calibration, pinned revision",". ",[32,1008,1013],{"href":1009,"ariaLabel":1010,"className":1011,"dataFootnoteBackref":37},"#user-content-fnref-threebit","Back to reference 1",[1012],"data-footnote-backref","↩"," ",[32,1016,1013,1020],{"href":1017,"ariaLabel":1018,"className":1019,"dataFootnoteBackref":37},"#user-content-fnref-threebit-2","Back to reference 1-2",[1012],[29,1021,47],{},[32,1023,1013,1027],{"href":1024,"ariaLabel":1025,"className":1026,"dataFootnoteBackref":37},"#user-content-fnref-threebit-3","Back to reference 1-3",[1012],[29,1028,62],{},[32,1030,1013,1034],{"href":1031,"ariaLabel":1032,"className":1033,"dataFootnoteBackref":37},"#user-content-fnref-threebit-4","Back to reference 1-4",[1012],[29,1035,213],{},[32,1037,1013,1041],{"href":1038,"ariaLabel":1039,"className":1040,"dataFootnoteBackref":37},"#user-content-fnref-threebit-5","Back to reference 1-5",[1012],[29,1042,230],{},[32,1044,1013,1048],{"href":1045,"ariaLabel":1046,"className":1047,"dataFootnoteBackref":37},"#user-content-fnref-threebit-6","Back to reference 1-6",[1012],[29,1049,725],{},[32,1051,1013,1055],{"href":1052,"ariaLabel":1053,"className":1054,"dataFootnoteBackref":37},"#user-content-fnref-threebit-7","Back to reference 1-7",[1012],[29,1056,1057],{},"7",[32,1059,1013,1063],{"href":1060,"ariaLabel":1061,"className":1062,"dataFootnoteBackref":37},"#user-content-fnref-threebit-8","Back to reference 1-8",[1012],[29,1064,1065],{},"8",[32,1067,1013,1071],{"href":1068,"ariaLabel":1069,"className":1070,"dataFootnoteBackref":37},"#user-content-fnref-threebit-9","Back to reference 1-9",[1012],[29,1072,1073],{},"9",[998,1075,1077,1006,1082,1014,1087,1014,1094,1014,1101],{"id":1076},"user-content-fn-twentyseven",[32,1078,1081],{"href":1079,"rel":1080},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FQwen3.8-27B-W3Rot-INT3-Paiton-RDNA4\u002Fblob\u002F383fa4b221a992fc2437500328056ef05355c8eb\u002FREADME.md",[907],"27B 3-bit model card and second accuracy update, pinned revision",[32,1083,1013],{"href":1084,"ariaLabel":1085,"className":1086,"dataFootnoteBackref":37},"#user-content-fnref-twentyseven","Back to reference 2",[1012],[32,1088,1013,1092],{"href":1089,"ariaLabel":1090,"className":1091,"dataFootnoteBackref":37},"#user-content-fnref-twentyseven-2","Back to reference 2-2",[1012],[29,1093,47],{},[32,1095,1013,1099],{"href":1096,"ariaLabel":1097,"className":1098,"dataFootnoteBackref":37},"#user-content-fnref-twentyseven-3","Back to reference 2-3",[1012],[29,1100,62],{},[32,1102,1013,1106],{"href":1103,"ariaLabel":1104,"className":1105,"dataFootnoteBackref":37},"#user-content-fnref-twentyseven-4","Back to reference 2-4",[1012],[29,1107,213],{},[998,1109,1111,1006,1116,1014,1121,1014,1128,1014,1135,1014,1142,1014,1149,1014,1156],{"id":1110},"user-content-fn-bench",[32,1112,1115],{"href":1113,"rel":1114},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002F4882e36e455c1e657175db910367c992837c72a5\u002Fmodels\u002FQwen3.8-Flash-Next\u002FBENCHMARKS.md",[907],"Flash Next performance measurements, cache validation and methodology, pinned release",[32,1117,1013],{"href":1118,"ariaLabel":1119,"className":1120,"dataFootnoteBackref":37},"#user-content-fnref-bench","Back to reference 3",[1012],[32,1122,1013,1126],{"href":1123,"ariaLabel":1124,"className":1125,"dataFootnoteBackref":37},"#user-content-fnref-bench-2","Back to reference 3-2",[1012],[29,1127,47],{},[32,1129,1013,1133],{"href":1130,"ariaLabel":1131,"className":1132,"dataFootnoteBackref":37},"#user-content-fnref-bench-3","Back to reference 3-3",[1012],[29,1134,62],{},[32,1136,1013,1140],{"href":1137,"ariaLabel":1138,"className":1139,"dataFootnoteBackref":37},"#user-content-fnref-bench-4","Back to reference 3-4",[1012],[29,1141,213],{},[32,1143,1013,1147],{"href":1144,"ariaLabel":1145,"className":1146,"dataFootnoteBackref":37},"#user-content-fnref-bench-5","Back to reference 3-5",[1012],[29,1148,230],{},[32,1150,1013,1154],{"href":1151,"ariaLabel":1152,"className":1153,"dataFootnoteBackref":37},"#user-content-fnref-bench-6","Back to reference 3-6",[1012],[29,1155,725],{},[32,1157,1013,1161],{"href":1158,"ariaLabel":1159,"className":1160,"dataFootnoteBackref":37},"#user-content-fnref-bench-7","Back to reference 3-7",[1012],[29,1162,1057],{},[998,1164,1166,1006,1171,1014,1176,1014,1183],{"id":1165},"user-content-fn-fourbit",[32,1167,1170],{"href":1168,"rel":1169},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FQwen3.8-Flash-Next-W4A8-Paiton-RDNA4\u002Fblob\u002F9ec8e306503d5f9c9e8fd9c0d77de8fa63428888\u002FREADME.md",[907],"Flash Next W4A8 model card and evaluation methods, pinned revision",[32,1172,1013],{"href":1173,"ariaLabel":1174,"className":1175,"dataFootnoteBackref":37},"#user-content-fnref-fourbit","Back to reference 4",[1012],[32,1177,1013,1181],{"href":1178,"ariaLabel":1179,"className":1180,"dataFootnoteBackref":37},"#user-content-fnref-fourbit-2","Back to reference 4-2",[1012],[29,1182,47],{},[32,1184,1013,1188],{"href":1185,"ariaLabel":1186,"className":1187,"dataFootnoteBackref":37},"#user-content-fnref-fourbit-3","Back to reference 4-3",[1012],[29,1189,62],{},[998,1191,1193,1006,1198,1014,1203,1014,1210,1014,1217,1014,1224,1014,1231,1014,1238,1014,1245],{"id":1192},"user-content-fn-release",[32,1194,1197],{"href":1195,"rel":1196},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002F4882e36e455c1e657175db910367c992837c72a5\u002Fmodels\u002FQwen3.8-Flash-Next\u002FREADME.md",[907],"Flash Next setup, cache modes, host requirements and native-worker limits, pinned release",[32,1199,1013],{"href":1200,"ariaLabel":1201,"className":1202,"dataFootnoteBackref":37},"#user-content-fnref-release","Back to reference 5",[1012],[32,1204,1013,1208],{"href":1205,"ariaLabel":1206,"className":1207,"dataFootnoteBackref":37},"#user-content-fnref-release-2","Back to reference 5-2",[1012],[29,1209,47],{},[32,1211,1013,1215],{"href":1212,"ariaLabel":1213,"className":1214,"dataFootnoteBackref":37},"#user-content-fnref-release-3","Back to reference 5-3",[1012],[29,1216,62],{},[32,1218,1013,1222],{"href":1219,"ariaLabel":1220,"className":1221,"dataFootnoteBackref":37},"#user-content-fnref-release-4","Back to reference 5-4",[1012],[29,1223,213],{},[32,1225,1013,1229],{"href":1226,"ariaLabel":1227,"className":1228,"dataFootnoteBackref":37},"#user-content-fnref-release-5","Back to reference 5-5",[1012],[29,1230,230],{},[32,1232,1013,1236],{"href":1233,"ariaLabel":1234,"className":1235,"dataFootnoteBackref":37},"#user-content-fnref-release-6","Back to reference 5-6",[1012],[29,1237,725],{},[32,1239,1013,1243],{"href":1240,"ariaLabel":1241,"className":1242,"dataFootnoteBackref":37},"#user-content-fnref-release-7","Back to reference 5-7",[1012],[29,1244,1057],{},[32,1246,1013,1250],{"href":1247,"ariaLabel":1248,"className":1249,"dataFootnoteBackref":37},"#user-content-fnref-release-8","Back to reference 5-8",[1012],[29,1251,1065],{},[998,1253,1255,1006,1260],{"id":1254},"user-content-fn-launcher",[32,1256,1259],{"href":1257,"rel":1258},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Fblob\u002F4882e36e455c1e657175db910367c992837c72a5\u002Fmodels\u002FQwen3.8-Flash-Next\u002Flaunch-flashnext.py",[907],"Flash Next launcher and pinned download selection",[32,1261,1013],{"href":1262,"ariaLabel":1263,"className":1264,"dataFootnoteBackref":37},"#user-content-fnref-launcher","Back to reference 6",[1012],[1266,1267,1268],"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 .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);}",{"title":37,"searchDepth":752,"depth":752,"links":1270},[1271,1275,1276,1277,1278,1279,1280,1281],{"id":66,"depth":752,"text":67,"children":1272},[1273,1274],{"id":234,"depth":762,"text":235},{"id":304,"depth":762,"text":305},{"id":325,"depth":752,"text":326},{"id":500,"depth":752,"text":501},{"id":618,"depth":752,"text":619},{"id":656,"depth":752,"text":657},{"id":692,"depth":752,"text":693},{"id":946,"depth":752,"text":947},{"id":36,"depth":752,"text":993},[977,1283,1284,1285,1286],"AMD Radeon","Local AI","Qwen","Quantization","2026-10-24T18:45:00Z","Paiton's new Qwen3.8 3-bit weights approach 16-bit task scores on two R9700 cards. Native MTP prefix caching resumes a 200K prompt in 0.68 seconds. Results, limits and setup.","md","3-bit accuracy. Faster conversation restarts.","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-3bit-mtp\u002Fhero.webp",{},"https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp","\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp",{"title":5,"description":1288},"paiton-qwen38-3bit-accuracy-prefix-caching-mtp","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-3bit-mtp\u002Fsocial.webp","blog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp",null,"WYHuHcddX9UKnaBpzq8opKNZblxD1AMOHiGqiI4CfHo",[1302,1304,1319,1327,1335,1343,1353,1366,1376,1388,1397,1412,1421,1433,1465,1476,1498,1516,1535,1553,1570,1587,1599,1615,1630,1642,1651,1659,1674,1686,1697,1708,1718,1731,1741,1754,1765,1775,1786,1795,1807,1818,1827],{"path":1294,"title":5,"description":1288,"date":1287,"slug":1296,"image":1291,"originalUrl":1293,"categories":1303},[977,1283,1284,1285,1286],{"path":1305,"title":1306,"description":1307,"date":1308,"slug":1309,"image":1310,"originalUrl":1311,"categories":1312},"\u002Fblog\u002Fnvidia-dsx-pod-vera-rubin-modular-data-centers","Nvidia confirms the modular DC direction","NVIDIA DSX POD brings Vera Rubin AI factories to modular data centers. ElioVP explains site readiness, cooling and the thinking behind ModFlex.","2026-10-21T13:00:00Z","nvidia-dsx-pod-vera-rubin-modular-data-centers","\u002Fasset\u002Fimages\u002Fblog\u002Fnvidia-dsx-pod-modflex\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fnvidia-dsx-pod-vera-rubin-modular-data-centers",[1313,1314,1315,1316,1317,1318],"Modular DC","AI Infrastructure","ModFlex","NVIDIA","Vera Rubin","Liquid Cooling",{"path":78,"title":1320,"description":1321,"date":1322,"slug":1323,"image":1324,"originalUrl":1325,"categories":1326},"Qwen3.8 Flash Next in 4-bit on 2 × Radeon R9700: tested against 16-bit","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.","2026-10-12T15:00:00Z","paiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700",[977,1283,1284,1285,1286],{"path":73,"title":1328,"description":1329,"date":1330,"slug":1331,"image":1332,"originalUrl":1333,"categories":1334},"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",[977,1283,1284,1285,1286],{"path":625,"title":1336,"description":1337,"date":1338,"slug":1339,"image":1340,"originalUrl":1341,"categories":1342},"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",[977,1283,1284,1285,1286],{"path":1344,"title":1345,"description":1346,"date":1347,"slug":1348,"image":1349,"originalUrl":1350,"categories":1351},"\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",[977,1283,1284,1352,1285],"Image Generation",{"path":1354,"title":1355,"description":1356,"date":1357,"slug":1358,"image":1359,"originalUrl":1360,"categories":1361},"\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",[977,1283,1362,1363,1364,1365],"vLLM","Qwen3.8","Inference Optimization","DFlash2",{"path":1367,"title":1368,"description":1369,"date":1370,"slug":1371,"image":1372,"originalUrl":1373,"categories":1374},"\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",[977,1283,1284,1375,1362],"GGUF",{"path":1377,"title":1378,"description":1379,"date":1380,"slug":1381,"image":1382,"originalUrl":1383,"categories":1384},"\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",[977,1283,1284,1385,1386,1387],"Video Generation","MiniMax H3","ComfyUI",{"path":1389,"title":1390,"description":1391,"date":1392,"slug":1393,"image":1394,"originalUrl":1389,"categories":1395},"\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",[977,1283,1284,1352,1396,1387],"FLUX",{"path":1398,"title":1399,"description":1400,"date":1401,"slug":1402,"image":1403,"originalUrl":1404,"categories":1405},"\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",[977,1406,1283,1407,1408,1409,1364,1410,1362,1411],"Artificial Intelligence","AI Inference","GPU Performance","Inference Latency","Large Language Models","Cost Efficiency",{"path":1413,"title":1414,"description":1415,"date":1416,"slug":1417,"image":1418,"originalUrl":1419,"categories":1420},"\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",[977,1406,1283,1407,1408,1409,1364,1410,1362,1411],{"path":1422,"title":1423,"description":1424,"date":1425,"slug":1426,"image":1427,"originalUrl":1299,"categories":1428},"\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",[1429,1314,1430,1315,1431,1432],"All","Data Centers","HPC","AMD Helios",{"path":1434,"title":1435,"description":1436,"date":1437,"slug":1438,"image":1439,"originalUrl":1440,"categories":1441},"\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",[1429,1406,977,1442,1443,1444,1445,1446,1447,1448,1449,1450,1451,1452,1453,1454,1455,1456,1457,1458,977,1459,1460,1461,1462,1463,1464],"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":1466,"title":1467,"description":1468,"date":1469,"slug":1470,"image":1471,"originalUrl":1472,"categories":1473},"\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",[1429,1406,1313,1474,1443,1475,1313,1456],"Uncategorized","De Tijd",{"path":1477,"title":1478,"description":1479,"date":1480,"slug":1481,"image":1482,"originalUrl":1483,"categories":1484},"\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",[1429,1406,1485,1474,1486,1487,1488,1489,1490,1491,1492,1493,1449,1494,1450,1495,1496,1497],"Trending","AI Act","Anthropomorphism","AVG","Benelux","ChatGPT","Cybersecurity","Data Governance","Data Security","GDPR","Microsoft Copilot","Privacy","Shadow AI",{"path":1499,"title":1500,"description":1501,"date":1502,"slug":1503,"image":1504,"originalUrl":1505,"categories":1506},"\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",[1429,1507,1508,1509,1491,1510,1511,1512,1494,1513,1514,1515,1496],"AWS","Belgian Mobile ID","Cloud Act","Data Sovereignty","Digital Identity","eIDAS","itsme","Liberty Global","MyGov.be",{"path":1517,"title":1518,"description":1519,"date":1520,"slug":1521,"image":1522,"originalUrl":1523,"categories":1524},"\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",[1429,1406,1525,1485,1526,1527,1528,1529,1530,1531,1532,1533,1459,1534],"Solutions","Agentic AI","AI Engineering","AI Strategy","Autonomous Agents","Enterprise AI","Local LLM","Model Fine-Tuning","On-Premise AI","VRAM Optimization",{"path":1536,"title":1537,"description":1538,"date":1539,"slug":1540,"image":1541,"originalUrl":1542,"categories":1543},"\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",[1429,1406,1485,1314,1544,1545,1546,1547,1548,1549,1550,1551,1552],"AI Neocloud","Circular Financing","GPU Cloud","Investment Risks","Startup Valuation","Synthetic Bubble","Tech Analysis","Vaporware","Venture Capital",{"path":1554,"title":1555,"description":1556,"date":1557,"slug":1558,"image":1559,"originalUrl":1560,"categories":1561},"\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",[1429,1313,1474,1562,1314,1563,1564,1318,1565,1566,1567,1568,1569],"150kW Rack","DLC","High Density","Modular Data Center","NVIDIA Blackwell Ultra","NVIDIA GB300","NVL72","Rapid Deployment",{"path":1571,"title":1572,"description":1573,"date":1574,"slug":1575,"image":1576,"originalUrl":1577,"categories":1578},"\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",[1429,1406,977,1474,1579,1406,1580,1581,1582,1583,1584,1585,977,1586],"AMD MI300X","CUDA Translation","FP8","GPU Optimization","High Performance Computing","HIP","Kernel Tuning","ROCm",{"path":1588,"title":1589,"description":1590,"date":1591,"slug":1592,"image":1593,"originalUrl":1594,"categories":1595},"\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",[1429,1406,977,1407,1596,1579,1411,1597,1364,1585,977,1598,1362],"AMD Instinct","High Throughput","SGLang",{"path":1600,"title":1601,"description":1602,"date":1603,"slug":1604,"image":1605,"originalUrl":1606,"categories":1607},"\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",[1429,1406,977,1608,1579,1609,1364,1610,1611,1612,1613,977,1614],"AI Benchmarks","Cost per Token","Mixture of Experts","MoE","NVIDIA B200","NVIDIA H200","Qwen3",{"path":1616,"title":1617,"description":1618,"date":1619,"slug":1620,"image":1621,"originalUrl":1622,"categories":1623},"\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",[1429,1406,1525,1474,1526,1624,1625,1626,1627,1531,1533,1459,1628,1629],"Damage Detection","Document Processing","Email Automation","Invoice Extraction","Ticket Automation","Workflow Automation",{"path":1631,"title":1632,"description":1633,"date":1634,"slug":1635,"image":1636,"originalUrl":1637,"categories":1638},"\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",[1429,1406,977,1474,1639,1443,1444,1640,1641,1455,1456,977,1362],"AI","H200","MI300X",{"path":1643,"title":1644,"description":1645,"date":1646,"slug":1647,"image":1648,"originalUrl":1649,"categories":1650},"\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",[1429,1406,1525,1526,1624,1625,1626,1627,1531,1533,1459,1628,1629],{"path":1652,"title":1653,"description":1654,"date":1655,"slug":1656,"image":37,"originalUrl":1657,"categories":1658},"\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",[1429,1406,977],{"path":1660,"title":1661,"description":1662,"date":1663,"slug":1664,"image":1665,"originalUrl":1666,"categories":1667},"\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",[1429,1406,977,1474,1407,1579,1668,1581,1669,1670,1671,977,1672,1673],"Cold Start","Graph Compilation","Llama 3.1 405B","LLM Optimization","Startup Latency","Tensor Parallelism",{"path":1675,"title":1676,"description":1677,"date":1678,"slug":1679,"image":1680,"originalUrl":1681,"categories":1682},"\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",[1429,1406,977,1474,1579,1683,1530,1450,1408,1409,1410,1670,1684,1685],"Cold Start Optimization","Model Serving","vLLM Optimization",{"path":1687,"title":1688,"description":1689,"date":1690,"slug":1691,"image":1692,"originalUrl":1693,"categories":1694},"\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",[1429,1406,977,1525,1474,1443,1641,1456,1695,1696],"RX7900XTX","tenstorrent",{"path":1698,"title":1699,"description":1700,"date":1701,"slug":1702,"image":1703,"originalUrl":1704,"categories":1705},"\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",[1429,1406,1313,1525,1444,1640,1706,1456,1707],"MI325x","pnl calculator",{"path":1709,"title":1710,"description":1711,"date":1712,"slug":1713,"image":1714,"originalUrl":1715,"categories":1716},"\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",[1429,1406,977,1639,1443,1640,1717,1456,977,1362],"MI300",{"path":1719,"title":1720,"description":1721,"date":1722,"slug":1723,"image":1724,"originalUrl":1725,"categories":1726},"\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",[1429,1313,1727,1314,1564,1431,1318,1565,1728,1729,1568,1730],"1-2MW Data Center","NVIDIA Blackwell","NVIDIA NVL","Precision Cooling",{"path":1732,"title":1733,"description":1734,"date":1735,"slug":1736,"image":1737,"originalUrl":1738,"categories":1739},"\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",[1429,1406,1525,1474,1639,1443,1740],"Healthcare",{"path":1742,"title":1743,"description":1744,"date":1745,"slug":1746,"image":1747,"originalUrl":1748,"categories":1749},"\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",[1429,1485,1639,1443,1750,1751,1752,1753],"import","Taiwan","Tariffs","Trump",{"path":1755,"title":1756,"description":1757,"date":1758,"slug":1759,"image":1760,"originalUrl":1761,"categories":1762},"\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",[1429,1406,1525,1639,1763,1764],"AI Agents","ERP",{"path":1766,"title":1767,"description":1768,"date":1769,"slug":1770,"image":1771,"originalUrl":1772,"categories":1773},"\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",[1429,1406,1485,1774,1443,1451,1456],"AI news",{"path":1776,"title":1777,"description":1778,"date":1779,"slug":1780,"image":1781,"originalUrl":1782,"categories":1783},"\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",[1429,1406,977,1639,1443,1784,1785,1641,977],"benchmark","LLM",{"path":1787,"title":1788,"description":1789,"date":1790,"slug":1791,"image":1792,"originalUrl":1793,"categories":1794},"\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",[1429,1406,977],{"path":1796,"title":1797,"description":1798,"date":1799,"slug":1800,"image":1801,"originalUrl":1802,"categories":1803},"\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",[1429,1443,1804,1805,1806],"Jim Greene","Podcast","Tech Talk",{"path":1808,"title":1809,"description":1810,"date":1811,"slug":1812,"image":1813,"originalUrl":1814,"categories":1815},"\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",[1429,1406,977,1443,1816,1817,1640,1641,1706,977,1362],"Deepseek","H100",{"path":1819,"title":1820,"description":1821,"date":1822,"slug":1823,"image":1824,"originalUrl":1825,"categories":1826},"\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",[1429,1406,977,1443,1816,1817,1640,1641,1706,977,1362],{"path":1828,"title":1829,"description":1830,"date":1831,"slug":1832,"image":1833,"originalUrl":1834,"categories":1835},"\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",[1429,1406,977,1443,1817,1640,1641,1706,977,1362],1792870331932]