The 26 September W3A4 benchmark with FP8 cache measured 492.1 aggregate output tok/s across eight requests and 19.9% faster weighted decode than the previous MXFP4 release. The newer image separately adds a 4-bit KV cache in the W3A4 65K mode; its capacity figure is not from the 492.1 tok/s run. In the original FP8-cache accuracy comparison, the 3-bit weights lowered the tested MMLU-Pro subset by 2.86 points; MXFP4 remains available.

Read the W3A4 benchmark Open the current release guide

Released image generation: Qwen-Image 2.1 v1.0.2

103.29 s

2048 × 2048 images on 1 × Radeon AI PRO R9700

about 37%less warm waiting vs the historical v1.0.1 repeat

25.55 GiBmaximum sampled whole-device VRAM

One 32 GB Radeon AI PRO R9700, 40 steps, guidance 1.0 and batch one. The 103.29 s median covers the warm HTTP request through PNG delivery across three published v1.0.2 containers. Startup and downloads are separate. The v1.0.1 comparison comes from a separate release repeat, not one matched experiment.

See the image benchmark Open the Qwen-Image setup guide

Video diffusion: MI355X vs B200

17.6%

Wan2.2-T2V-A14B ran faster on AMD MI355X

5.501sAMD MI355X

6.672sNVIDIA B200

Paiton-Diffusers removed runtime waste from the generation path and let the hardware show up.

Read Wan2.2 Benchmark

Local video + audio: Radeon AI PRO R9700

16.7%

Less waiting for MiniMax H3 video with sound

5m 33smean time to a saved MP4

15svideo with native stereo audio

One 32 GB R9700 generates 864 × 480 video with native stereo audio, fully local after setup. Compared with matched stock Turbo8, one prompt and two measured requests per engine after warmup. Timing includes saving the MP4.

Watch the clips

Why it matters

You do not need a new model to improve unit economics. You need the runtime, kernels, and deployment path tuned around the workload you already run.

[ 00 / 10 ]

[ GPU ECONOMICS MODEL ]

TRANSLATE THROUGHPUT INTO REUSABLE CAPACITY.

EDITABLE PLANNING SCENARIO

Turn runtime gains
into a business case.

Faster inference is more than a benchmark number. It can reduce the GPU time required for the same workload,

or increase the output of hardware you already own.

Start with this illustrative scenario, then replace the assumptions with the uplift measured during your Paiton benchmark.

GPUs in scopeGPUsActive hours per GPU/monthFully loaded GPU cost€/hourExpected or measured gain%

SAME WORKLOAD13.0% fewer GPU-hours per unit of output

16,000 current GPU-hours / month

Current runtime baselinePaiton-adjusted estimate: 13,913 GPU-hours

CAPACITY RELEASED

2,087GPU-hours / month

Equivalent to roughly 4.2 GPUs at this utilization level.

Potential compute-capacity value€6,261 / month€75,130 / year

SAME COMPUTE BUDGET

+15.0%potential output

Use the released runtime for more tokens, generations, jobs, or customer demand without expanding the GPU budget.

A published Radeon AI PRO R9700 test measured 39.77 versus 32.86 output tokens per second on the same Qwen3.8 workload. Gains remain workload-specific and that result is not applied automatically here.

View the R9700 benchmark

Benchmark my workload

Illustrative planning model. Actual gains depend on the model, batch size, sequence length, precision, runtime, hardware, scaling behaviour, and deployment. The calculation assumes throughput improvement translates proportionally into lower GPU runtime for the same workload. Released capacity may become a cash saving in usage-based infrastructure or reusable capacity on owned hardware. Figures exclude Paiton fees and do not guarantee savings or performance.

[ 00 / 10 ]

[ WHAT PAITON CHANGES ]

LANGUAGE, IMAGES, VIDEO WITH SOUND, AND CUSTOM STACKS.

Turn AMD hardware
into production throughput.

Whether you run LLMs, image and video diffusion, MoE, or custom architectures,
Paiton starts where money is lost: latency, throughput, memory pressure, and cost per useful output.

More tokens, less waiting

Serve more work from one GPU. In the 26 September Qwen3.8 W3A4 65K benchmark, one R9700 reached 492.1 aggregate output tok/s across eight requests. This is combined throughput, not the speed of one response; the newer 4-bit KV profile was tested separately.

No retraining required

Keep the checkpoint where appropriate, or choose calibrated lower-bit weights when quality checks justify them. Neither path requires retraining the base model.

Runtime work that pays back

ComfyUI workflows for FLUX.2 klein images and MiniMax H3 video with sound on R9700, a separate Qwen-Image 2.1 image API, Wan2.2 video on Instinct, and vLLM for language inference.

Language Models

Llama, Qwen, Deepseek, Gemma, Mistral, and CodeLlama with vLLM support, FP8 precision, and custom kernels when the gain is worth it.

Diffusion & Video

Qwen-Image 2.1, MiniMax H3 with native audio, Wan2.2, FLUX.2 klein, Stable Diffusion, SDXL and ControlNet, with workload-specific runtime and memory tuning.

Advanced Models

MoE, MLA, multimodal systems, proprietary architectures, expert kernels, novel attention, and custom inference stacks.

[ 00 / 10 ]

[ ENGAGEMENT FLOW ]

MEASURE. TUNE. PROVE.

A practical optimization sprint

Bring the workload, the target GPU, and the performance goal.
We turn that into a focused path from measurement to production-ready gain.

01Measure

Run the real workload and identify where latency, memory pressure, or cost per useful output is leaking value.

02Tune

Build the AMD-specific path with custom .so files, kernel fusion and a precision choice validated against the workload.

03Prove

Deploy the optimized path on AMD Instinct or Radeon AI PRO hardware and compare the result against the baseline.

[ 00 / 10 ]

[ DEPLOYMENT FIT ]

ROCM, KERNELS, MULTI-GPU, AND RUNTIME WORK.

Serious AMD tuning,
for production inference.

Runtime work

vLLMPaiton-DiffusersComfyUI (FLUX.2 klein and MiniMax H3 on R9700)Qwen-Image 2.1 image API (R9700)ROCmHIPIndependent kernels

Performance levers

FP8 precisionQualified 3-bit weights and 4-bit KVTensor parallelismData parallelismKernel fusionCustom AMD operatorsMulti-GPU scaling

What gets tuned

Precision is a workload choice. In the measured Qwen3.8 profile, W3A4 improves speed and cache headroom with a knowledge-accuracy tradeoff versus MXFP4; other models need their own checks. Tensor parallelism spreads large models like Llama-3.1-405B across multiple AMD GPUs. Custom kernels target the bottlenecks generic runtimes leave behind.

CDNA 4.0

MI355X

288GB HBM3E

CDNA 3.0

MI325X / MI300X / MI300A

256GB HBM3E, 192GB HBM3, 128GB HBM3 APU

CDNA 2.0

MI250X / MI250 / MI210

Datacenter AMD GPU support

CDNA 1.0

MI100

First compute-optimized generation

RDNA 4.0

Radeon AI PRO R9700

Qualified Qwen3.8, Ornith 1.5, FLUX.2 klein, Qwen-Image 2.1 and MiniMax H3 profiles on 32GB GDDR6

RDNA 3.0

RX 7900 XTX / RX 7900 XT

Supported for selected inference workloads

RDNA 2.0

RX 6800 XT / RX 6900 XT

Supported with workload-specific qualification

Paiton supports both AMD Instinct datacenter accelerators and Radeon GPUs. Every model, runtime, precision, and hardware combination is qualified against a defined deployment scope before delivery.

[ 00 / 10 ]

[ COMMUNITY ]

PRACTICAL LOCAL AI, SHARED WITH THE COMMUNITY.

Faster local AI.Shared with the community.

Better inference should reach beyond our customer projects. Alongside our open-source vLLM plugin, our public RDNA community packages help developers run language models, generate images and create video with sound on their own hardware.

Explore the code, run the supported models locally, and share what you learn. It is our way of helping more people get more from the GPUs they already own.

Plan your on-prem AI

Community packagesvLLM plugin: Apache-2.0

paiton-vllm-plugin

vLLM integration, a unified Qwen3.8 Docker image with 65K and 200K modes, optional W3A4 weights and 65K image input, a separate Qwen-Image 2.1 image API, ComfyUI workflows and setup guides.

Explore the packages on GitHub

Qwen3.8 supports structured tool calls. The current R9700 Docker image defaults to 65K text mode; optional W3A4 weights enable its 4-bit KV cache there. The earlier W3A4/FP8-cache accuracy comparison found a knowledge-accuracy tradeoff versus MXFP4; the 4-bit KV setting was checked separately. The 200K mode uses one active request, FP8 KV and prefix caching; 220K was tested separately. Optional image input is available in 65K mode and reduces cache capacity. These Docker profiles differ from the native paiton serve qwen38-nvfp4 preset. Separate R9700 packages cover Qwen-Image 2.1 and FLUX.2 klein images plus MiniMax H3 video with sound. Qwen-Image weights use the Qwen Research License for research and evaluation; commercial use needs a separate upstream license. The vLLM plugin is Apache-2.0 and the Paiton compiler remains private.

[ 00 / 10 ]

[ EVIDENCE ]

PUBLISHED BENCHMARKS AND CASE STUDIES.

Read the benchmarks
behind the claims

492.1 aggregate tok/s

Qwen3.8 W3A4: faster decode on one R9700

The 26 September 65K test used FP8 KV cache: eight concurrent requests reached 492.1 aggregate output tok/s and weighted decode rose 19.9% to 184.4 tok/s versus the previous MXFP4 release. The later 4-bit KV and 200K profiles were measured separately. In the original FP8-cache accuracy check, W3A4 reduced the tested MMLU-Pro subset score by 2.86 points.

Read the W3A4 benchmark

103.29 s warm image request

Qwen-Image 2.1 on one Radeon AI PRO R9700

The published v1.0.2 default created a 2048 × 2048 PNG in a 103.29 s median warm request across three containers. Same MXFP4 checkpoint, 40 steps, guidance 1.0 and batch one. About 37% lower warm latency than the separate historical v1.0.1 repeat; quality was checked against the exact profile.

Read the image benchmark

400.7 aggregate tok/s

Earlier Qwen3.8 MXFP4 result on ROCm 10

The earlier 65K MXFP4 preset reached 400.7 aggregate tok/s at eight concurrent requests and 146.9 weighted serial decode tok/s. Its separate article covers long-context chat and preserves the original 57% gain against matched Radiance + DFlash2 in the 188-request, 8K-context comparison.

Benchmark + demo

16.7% less waiting

MiniMax H3 video and audio on one Radeon

A 15-second clip with native stereo audio, saved in 5m 33s on average versus 6m 39s for matched stock. Turbo8 on one 32 GB R9700, with warm, complete-request timings.

Watch the clips

16.2% less generation time

FLUX.2 klein on Radeon AI PRO R9700

1.054 seconds per image with a warm pipeline, 33.4% lower peak Torch allocation, and a free local ComfyUI workflow. Tested at 1024 × 1024, four steps and batch one.

Read Benchmark

+27% output

Ornith 1.5 on one 32GB Radeon GPU

Paiton with DFlash delivers 44.63 output tokens per second on R9700, 27% ahead of tuned stock vLLM in the tested single-user workload.

Read Benchmark

+21% output

Qwen3.8 on Radeon AI PRO R9700

Higher interactive, coding, and long-context throughput from the same 32GB Radeon GPU.

Read Benchmark

405B model

Llama-3.1-405B acceleration

Faster startup and inference for a frontier-scale open model.

Read Case Study

$/1M tokens

MoE kernels beat H200/B200

MI300X plus Paiton runtime on long-context MoE economics.

Read Benchmark

[ 00 / 10 ]

[ NEXT STEP ]

START WITH THE MODEL AND TARGET GPU.

Show us the workload.
We will show you the gain.

Share the model, target GPU, and production goal.
Paiton turns that into a benchmark-led optimization plan.

Request a workload benchmark

Get more performance from AMD GPUs.

measured gains

Benchmarks that make AMD worth a second look

Faster text inference, with a precision choice

2048 × 2048 images on 1 × Radeon AI PRO R9700

Wan2.2-T2V-A14B ran faster on AMD MI355X

Less waiting for MiniMax H3 video with sound

Turn runtime gains into a business case.

Turn AMD hardware into production throughput.

A practical optimization sprint

01Measure

02Tune

03Prove

ROCM, KERNELS, MULTI-GPU, AND RUNTIME WORK.

Serious AMD tuning, for production inference.

Runtime work

Performance levers

What gets tuned

Faster local AI.Shared with the community.

paiton-vllm-plugin

PUBLISHED BENCHMARKS AND CASE STUDIES.

Read the benchmarks behind the claims

Qwen3.8 W3A4: faster decode on one R9700

Qwen-Image 2.1 on one Radeon AI PRO R9700

Earlier Qwen3.8 MXFP4 result on ROCm 10

MiniMax H3 video and audio on one Radeon

FLUX.2 klein on Radeon AI PRO R9700

Ornith 1.5 on one 32GB Radeon GPU

Qwen3.8 on Radeon AI PRO R9700

Llama-3.1-405B acceleration

MoE kernels beat H200/B200

START WITH THE MODEL AND TARGET GPU.

Show us the workload. We will show you the gain.

Get more performance
from AMD GPUs.

Benchmarks that make AMD
worth a second look

Turn runtime gains
into a business case.

Turn AMD hardware
into production throughput.

Serious AMD tuning,
for production inference.

Read the benchmarks
behind the claims

Show us the workload.
We will show you the gain.