Builds / AMD Ryzen AI Max+ 388 (Radeon 8060S iGPU)
Ryzen AI Max+ 388 mini PC (64 GB): Qwen3.6 35B-A3B at 68.0 tok/s, still 21.7 at 256K context (llama.cpp ROCm, GEEKOM records)
Last updated:
- GPU(s)
- AMD Ryzen AI Max+ 388 (Radeon 8060S iGPU)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 64
- CPU
- AMD Ryzen AI Max+ 388 (8 cores / 16 threads)
Amazon · eBay (used) · Newegg - Motherboard
- GEEKOM A9 Mega 388 mini PC
Amazon · eBay (used) · Newegg - RAM
- 64 GB unified memory (256 GB/s)
- PSU
- not stated
- Other
- Ubuntu 26.04, ROCm 10.0.0, graphics driver 31.50.0.0; one request at a time (llama-server -np 1 or default)
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.6 35B-A3B | UD-Q4_K_M (22.7 GB) + MTP draft N=2 | llama.cpp b11065 (ROCm) | 68.03 | 1,240 @ 8,192-token prompt est. TTFT @8k ≈ 6.6 s (calculated from measured prefill) measured TTFT: 6.6 s @ 8K-token prompt | ⚡ spec-decode (MTP)context: 128-token input📏 Decode rate from the 128-token-input record (3,094 tokens generated); 67.76 at 8K, 54.3 at 32K input |
| Qwen3.6 35B-A3B | UD-Q4_K_M (22.7 GB) + MTP draft N=2 | llama.cpp b11065 (ROCm) | 21.66 | not stated | ⚡ spec-decode (MTP)context: 256K input📏 Decode at a 256K-token input record; TTFT 645 s at that length (prefill 406.4 tok/s) |
| Qwen3.8 27B | UD-Q4_K_XL (17.6 GB) + MTP draft N=4, KV q8_0 | llama.cpp (ROCm) | 16.87 | 299 @ 32,768-token prompt est. TTFT @8k ≈ 27.4 s (calculated from measured prefill) | ⚡ spec-decode (MTP)context: 32K input📏 Decode at a 32K input record (20.24 at 128-token input, 6.06 at 256K); listed at 32K as the sustained figure |
| Gemma 4 31B | q4_0 | llama.cpp (ROCm) | 11.28 | 69.53 @ 8,192-token prompt est. TTFT @8k ≈ 118 s (calculated from measured prefill) | context: 128-token input📏 Decode at 128-token input, no speculative decoding; 8.56 at 32K and 5.0 at 128K input (128K TTFT 10,312 s); the 256K run failed |
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.
Vendor lab measurements (GEEKOM tests its own A9 Mega; disclosed by the vendor), published 24 Oct 2026 with every record and launch command (CC BY 4.0 dataset). Qwen rows use MTP speculative decoding, so they are not comparable with plain decode. The 64 GB 388 cannot load Qwen3.8 Flash-Next alone; split across two 388 units over llama.cpp RPC it decoded at 30.76 tok/s (128-token input), with no speed gain from a third node. Sub-Q4 DeepSeek V4 Flash (UD-Q2_K_XL) runs from the same set are not listed.
https://huggingface.co/blog/geekom-ai-admin/measured-not-claimed-72-local-llm-benchmark-record
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/geekom-a9-mega-ryzen-ai-max-388-64gb-qwen36-qwen38-gemma4/)
HTML
<a href="https://getrigbase.com/builds/geekom-a9-mega-ryzen-ai-max-388-64gb-qwen36-qwen38-gemma4/" rel="noopener"><img src="https://getrigbase.com/badge/geekom-a9-mega-ryzen-ai-max-388-64gb-qwen36-qwen38-gemma4.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can Strix Halo (Radeon 8060S) run Gemma 4 31B? · Can Strix Halo (Radeon 8060S) run Qwen3.6 35B-A3B? · Can Strix Halo (Radeon 8060S) run Qwen3.8 27B?