Builds / AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU, gfx1151)
Strix Halo (Ryzen AI Max+ 395, 128 GB): Qwen3.8 Flash-Next UD-IQ4_XS, 16.8 tok/s plain, 25-32 tok/s with MTP on ROCm
Last updated:
Source: github.com ↗More source: github.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
- GPU(s)
- AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU, gfx1151)
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
- ROCm 7.1; patched llama.cpp branch drluoto/strix-halo-flash-next (hipCUB / radix TOP_K fix + native MTP draft head + ngram-mod)
- 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 | UD-IQ4_XS, no speculation | llama.cpp ROCm (drluoto/strix-halo-flash-next branch) | 16.8 | not stated | context: 8K ctx, coding workload (file rewrite / new code)📏 Single-stream greedy decode as printed in the source table; ~15 tok/s at 24K ctx |
| Qwen3.8 Flash-Next | UD-IQ4_XS + native MTP draft head + ngram-mod | llama.cpp ROCm (drluoto/strix-halo-flash-next branch), --spec-type draft-mtp,ngram-mod | 31.7 | not stated | ⚡ spec-decode (MTP)context: 8K ctx, "new code" task📏 Lower of the two 8K coding tasks (47.1 on a file rewrite, which repeats input and favours n-gram drafting) |
| Qwen3.8 Flash-Next | UD-IQ4_XS + native MTP draft head + ngram-mod | llama.cpp ROCm (drluoto/strix-halo-flash-next branch), --spec-type draft-mtp,ngram-mod | 25.4 | not stated | ⚡ spec-decode (MTP)context: 24K ctx, "new code" task📏 Lowest number in the source table (28.6 on a file rewrite at 24K) |
drluoto, llama.cpp discussion #27950, 29 Aug 2026. Coding workloads, greedy, single stream. The author checked the output byte-clean at long prompts, warning that an earlier community MTP port showed good tok/s while emitting multilingual noise above ~1k prompt tokens.
Source: https://github.com/ggml-org/llama.cpp/discussions/27950
https://github.com/drluoto/flash-next-strix-halo
https://github.com/drluoto/flash-next-strix-halo
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Embed this result
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-rocm-mtp-drluoto/)
HTML
<a href="https://getrigbase.com/builds/strix-halo-128gb-qwen38-flash-next-rocm-mtp-drluoto/" rel="noopener"><img src="https://getrigbase.com/badge/strix-halo-128gb-qwen38-flash-next-rocm-mtp-drluoto.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)
Quick answers: Can Strix Halo (Radeon 8060S) run Qwen3.8 Flash-Next?