Builds / AMD Ryzen AI Max+ 395 (Radeon 8060S iGPU)
Ryzen AI Max+ 395 (128 GB): Qwen3.8 27B GSQ IQ3_S on llama.cpp Vulkan at 16.1 tok/s, 28.8 with MTP (Kingy)
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 (shared memory)
- PSU
- not stated
- Other
- Ubuntu 26.04.1, llama.cpp build 79e2e74eb (Vulkan), whole model in graphics memory
- 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 27B | GSQ-RCO IQ3_S + MTP head (ISTA-DASLab GGUF, 12.12 GB) | llama.cpp 79e2e74eb (Vulkan) | 16.1 | not stated | context: 32K ctx, one request at a time📏 Kingy test 8 Nov 2026: llama-server, 32K ctx, prompt cache off, temperature 0, seed 42, up to 1,024 reply tokens (all spent thinking); average of 4 prompts (code, explainer, HTML report, question on an 8,217-token document); MTP off; llama-bench tg128 16.7 (16.0 with 8K already in context); about 25-26 s before the first token on the 8K document |
| Qwen3.8 27B | GSQ-RCO IQ3_S + MTP head (ISTA-DASLab GGUF, 12.12 GB) | llama.cpp 79e2e74eb (Vulkan) | 28.8 | not stated | ⚡ spec-decode (MTP)context: 32K ctx, one request at a time📏 Kingy test 8 Nov 2026: llama-server, 32K ctx, prompt cache off, temperature 0, seed 42, up to 1,024 reply tokens (all spent thinking); average of 4 prompts (code, explainer, HTML report, question on an 8,217-token document); --spec-type draft-mtp --spec-draft-n-max 2, 71% of drafts kept; 3 or 4 drafts were slower (27.8 and 27.4) |
| Qwen3.8 27B | UD-Q3_K_XL (Unsloth) | llama.cpp (Vulkan) | 15.8 | not stated | context: llama-bench📏 llama-bench tg128, Kingy's earlier run of 6 Nov 2026 on the same machine, quoted in the same article |
Owner-run test by Kingy (published 8 Nov 2026). The file is ISTA-DASLab's GSQ-RCO IQ3_S (3.5 bpw) plus the model's MTP head; every copy matched the published SHA-256. Small samples (most speed settings ran once or twice). The same file scored 20/20 on Kingy's 20-task quality test (run on a MacBook Pro). Same article: MacBook Pro M4 Pro and RTX 5070 Laptop builds.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Quick answers: Can Strix Halo (Radeon 8060S) run Qwen3.8 27B?
Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).