AMD local LLM tests: the configuration behind the score
Recent AMD results make a broad claim such as “Radeon is slow” or “ROCm is fixed” difficult to defend. The specific engine, model build and deployment environment can change the outcome substantially.
Research checked 31 October 2026. Compiled from published tests. No hardware benchmarks were run for this article; measurements belong to the credited testers.
Donato Capitella, who publishes as kyuz0 and maintains Strix Halo Toolboxes, now separates speed measurements from task-completion tests. His current speed site includes October 29 measurements for four Radeon AI PRO R9700 cards running Qwen3.8-27B with Standalone Radiance and ROCm 10.1.0. The displayed 64K-context, single-request FP8 rows report generation rates of 63.78 tokens per second without speculative decoding and 133.89 with the dflash2-7 configuration. Both specify tensor parallelism across four cards and an FP8 attention cache. Capitella’s current speed results.
That result describes two documented configurations on a four-card workstation. It is not the expected speed of one R9700, and it does not establish output quality or performance on an unrelated model. Speculative decoding uses a draft process to propose tokens for verification, so the draft configuration belongs beside the score.
A corrected result is more relevant than an old headline
Dustin Moore’s Puget Systems article contains two September corrections worth reading before quoting its results. A tokenizer mismatch inflated an earlier DeepSeek-R1-Distill-Llama-8B throughput figure; the corrected single-user result is 34.5 tokens per second. Puget also retracted an initial stock-vLLM multi-GPU failure claim after finding the problem in its virtualized test environment and successfully retesting on bare metal. Puget Systems’ corrections.
The corrected Qwen3.6-27B test reports 15.88 tokens per second for one user on two R9700s with stock vLLM FP8 tensor parallelism, and 156.21 aggregate tokens per second at eight concurrent users. That second number describes shared throughput, not the speed each person sees. Puget sells workstations, which is relevant commercial context for its testing.
Strix Halo has its own tests
Capitella’s Terminal Bench Mini dataset, generated October 27, includes a Strix Halo run of Qwen3.8-Flash-Next UD-Q4_K_XL using Gufo, ROCm 10.0 and an mtp-7 profile. It records 19 of 19 tasks passed on the first attempt. The benchmark is a custom 19-task subset with long time allowances; the result should not be generalized to all coding work or presented as a hardware comparison against differently configured runs. Terminal Bench Mini.
AMD’s October 5 ROCm 10.1 announcement supplies the release context, including installation tooling and WSL2 support in tech preview. It does not validate every Radeon card or Windows workflow. AMD release article.
For readers evaluating AMD, the relevant question is whether a recent, documented configuration exists for their intended workload. These sources provide several. They also show why a result without the engine version and operating environment leaves out too much.
Read the series
- Mac Studio and RTX cards: reading published LLM tests
- DGX Spark 64GB: what published tests can tell us
- AMD local LLM tests: the configuration behind the score (this article)
- Why local LLM benchmark charts disagree
- Local AI cost calculations: power, usage and payback
For memory planning, use the model-fit guide. Browse all Local AI notes.
