Builds / AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU)
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 Flash-Next at up to 1M context on a Vulkan fork, 31-39 tok/s at 32k (vahpetr)
Last updated:
- GPU(s)
- AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 128
- CPU
- AMD Ryzen AI Max+ 395
Amazon · eBay (used) · Newegg - Motherboard
- not stated
- RAM
- 128 GB unified memory
- PSU
- not stated
- Other
- Vulkan/RADV; llama.cpp fork vahpetr/llama.cpp-s1-fork (branch s1-public), KV q8_0, one slot. Model is an Uncensored fine-tune of Qwen3.8 Flash-Next
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.8 Flash-Next | i1-Q4_K_M (Uncensored fine-tune) + MTP draft | llama.cpp Vulkan (s1 fork) | 38.7 | 450 @ 32,768-token prompt est. TTFT @8k ≈ 18.2 s (calculated from measured prefill) | ⚡ spec-decode (MTP)context: 32k context depth📏 Worst of 3 measurement windows from the repo's BENCHMARKS.md, 25.83 ms/token converted to tok/s |
| Qwen3.8 Flash-Next | i1-Q4_K_M (Uncensored fine-tune), no draft | llama.cpp Vulkan (s1 fork) | 30.6 | not stated | context: 32k context depth📏 Serial decode without the draft, 32.69 ms/token from BENCHMARKS.md converted to tok/s |
| Qwen3.8 Flash-Next | i1-Q4_K_M (Uncensored fine-tune) + MTP draft | llama.cpp Vulkan (s1 fork) | 35.5 | not stated | ⚡ spec-decode (MTP)context: 1M context depth📏 Decode at a 1M-token context as posted in the discussion (61 tok/s on an empty context, 45 at 524k) |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
vahpetr, llama.cpp discussion #30282 (10 Oct 2026, updated 16 and 21 Oct). The author reports stock llama.cpp at about 24-27 tok/s on the same box. We list the conservative sustained figures, not the 61 tok/s empty-context peak. This is a community fork that is not merged upstream, and the weights are a third-party Uncensored fine-tune. The no-draft decode at 1M context was 24.6 tok/s (40.65 ms/token).
https://github.com/vahpetr/llama.cpp-s1-fork
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.
Markdown
[](https://getrigbase.com/builds/strix-halo-128gb-qwen38-flash-next-1m-context-vulkan-s1-vahpetr/)
HTML
<a href="https://getrigbase.com/builds/strix-halo-128gb-qwen38-flash-next-1m-context-vulkan-s1-vahpetr/" rel="noopener"><img src="https://getrigbase.com/badge/strix-halo-128gb-qwen38-flash-next-1m-context-vulkan-s1-vahpetr.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can Strix Halo (Radeon 8060S) run Qwen3.8 Flash-Next?