512 concurrent batch requests
DeepSeek V4.1 Flash on 1 × AMD Instinct MI350P
Paiton reached 77.9 single-request tok/s in LocalMaxxing. A separate batch profile served 512 simultaneous requests, reaching 1,107.6 aggregate tok/s with 8K-token outputs. One MI350P at a 450 W GPU limit, about 290 GB of host memory and CPU expert computation. Batch throughput is not interactive chat speed. The native CDNA4 runtime is proprietary.
Read the DeepSeek benchmark273.1 single-request tok/s
Qwen3.8 Flash Next on 1 × AMD Instinct MI350P
At a configured 450 W, Paiton reached 273.1 single-request output tok/s and 2,508.6 aggregate tok/s at 64 concurrent requests in LocalMaxxing. The comparison uses Paiton A5 weights versus AMD Quark MXFP4 on vLLM and SGLang, so it measures complete serving stacks, not the engine alone. Includes request-latency sweeps, GPU-only energy measurements and downloadable evidence. The optimized CDNA4 engine is proprietary.
Explore the MI350P benchmark258.7 weighted tok/s
Better 3-bit accuracy, faster conversation restarts
On two R9700 cards, the 24 October native-worker benchmark reports 258.7 weighted single-request tok/s. Revised 3-bit weights show no statistically significant gap from the 16-bit reference on the tested math, code and knowledge tasks; this does not establish equivalence, and vision still trails. Optional native prefix caching with MTP reuses long conversations; cache mode and host-memory limits apply.
Read the 24 October update492.1 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 benchmark103.29 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 benchmark400.7 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 + demo16.7% 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 clips16.2% 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 Benchmark405B 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
