What does @c4software run on Strix Halo?

Two 128 GB AMD Ryzen AI Max+ 395 boxes. bigchuck runs llama.cpp in a ROCm container with Qwen3.8-27B, Qwen3.8-Flash-Next and DeepSeek V4 Flash, mostly at UD-Q4_K_XL (DeepSeek on an IQ2_XXS mix), with speculative decoding where it pays. silentchuck is a Minisforum MS-S1 MAX running gufo, his Strix-specific HIP engine, on the same models. Self-reported, 24 October 2026.

How should a Strix Halo speed figure be labelled?

Engine build, quant, context length, the prompt, whether speculative decoding was on, several runs with the spread, and a correctness check. A broken backend can post a high tokens-per-second figure and still write garbage.

@c4software's two 128 GB Strix Halo boxes

Two 128 GB AMD Ryzen AI Max+ 395 computers. bigchuck runs llama.cpp. silentchuck, a Minisforum MS-S1 MAX, runs his Strix-specific engine gufo. The speed figures below are separate gufo versions, not one run.

@c4software. The handle is the credit he approved for publication. 24 Oct 2026. Self-reported. Not a localLLMrig measurement.

Test Gemma 4 31B, Recursive Flint and Nemotron 3.5 Super VL side by side in your browser, free.

GitHub: github.com/c4software/llmsetup. Current gufo numbers: docs/GUFO.md.

The two boxes

Two Strix Halo boxes, both with 128 GB. bigchuck runs llama.cpp in a ROCm container: Qwen3.8-27B, Qwen3.8-Flash-Next, DeepSeek V4 Flash, and smaller models. The quants are mostly UD-Q4_K_XL. DeepSeek uses an IQ2_XXS mix. Speculative decoding (MTP, DFlash, DSpark, or n-gram) is on where it pays.

silentchuck is a Minisforum MS-S1 MAX with 128 GB. It runs gufo, his Strix-specific HIP engine, on the same models, and he re-measures every release. Current gufo numbers are in docs/GUFO.md in the llmsetup repository.

Agentic decode of 50.7 tokens per second was gufo 0.7.1.

Prefill of 1,424 tokens per second at 52,000 tokens of context was gufo 0.8.1.

Those are different gufo versions, not one run. Since gufo 0.11.0 his gufo measurements run on silentchuck, a Minisforum MS-S1 MAX with a Ryzen AI Max+ 395 and 128 GB, not on bigchuck.

The next upgrade

A dedicated graphics card, mainly for prefill. Long agentic context is where unified memory hurts. Decode is fine. The budget is under €3,000, and it is not in the current timeline.

New model hardware requirements and price updates by email. We email a link to confirm the address. Nothing is sent until you click it. Privacy notice.

Where he looks before buying

GitHub issues and pull requests, llama.cpp discussions, model cards, and Discord. What is missing on a speed figure: the engine build, the context length, the prompt, and whether speculative decoding was on or off.

What would make a speed figure trustworthy

A published method and the raw logs: build, quant, context, prompt, and several runs with the spread between them. Plus a correctness check on the output. A broken backend can produce excellent tokens per second while generating garbage.

What he wishes a hardware site told him

Real-workload numbers: long-context prefill, prompt-cache reuse, and wall time on a real agentic loop. Not short-prompt decode on its own.

On our list: Can I run Qwen3.8-27B?, Can I run Qwen3.8-Flash-Next?, Can I run DeepSeek V4.1 Flash 1017?, GMKtec EVO-X2 128GB.

The numbers on this page were measured by the owner on that hardware. They are not a localLLMrig benchmark, and they do not change the fit scores in the calculator.

Questions

What does @c4software run on Strix Halo?: Two 128 GB AMD Ryzen AI Max+ 395 boxes. bigchuck runs llama.cpp in a ROCm container with Qwen3.8-27B, Qwen3.8-Flash-Next and DeepSeek V4 Flash, mostly at UD-Q4_K_XL (DeepSeek on an IQ2_XXS mix), with speculative decoding where it pays. silentchuck is a Minisforum MS-S1 MAX running gufo, his Strix-specific HIP engine, on the same models. Self-reported, 24 October 2026.
Were the gufo 0.7.1 decode and the gufo 0.8.1 prefill one run?: Agentic decode of 50.7 tokens per second was gufo 0.7.1. Prefill of 1,424 tokens per second at 52,000 tokens of context was gufo 0.8.1. Those are different gufo versions, not one run. Since gufo 0.11.0 his gufo measurements run on silentchuck, a Minisforum MS-S1 MAX with a Ryzen AI Max+ 395 and 128 GB, not on bigchuck. Current numbers are in docs/GUFO.md.
Which computer are the current gufo numbers measured on?: Since gufo 0.11.0 his gufo measurements run on silentchuck, a Minisforum MS-S1 MAX with a Ryzen AI Max+ 395 and 128 GB, not on bigchuck.
What is @c4software buying next?: A dedicated graphics card, mainly for prefill, because long agentic context is where unified memory hurts. Decode is fine. Budget under €3,000, not in the current timeline.
How should a Strix Halo speed figure be labelled?: Engine build, quant, context length, the prompt, whether speculative decoding was on, several runs with the spread, and a correctness check. A broken backend can post a high tokens-per-second figure and still write garbage.

Find hardware · All builder interviews · Send us your llama-bench result