Independent guide · methods and evidence
Qwen3.8 27B on AMD Strix Halo
Run Qwen3.8 27B on AMD Strix Halo: measured Ollama performance, image and tool checks, context limits, and separately labelled experimental routes.
Evidence reviewed .
September 19 functional update: the scoped runtime qualification records text (Qwen3.6), image (Qwen2.5-VL), executed tools (Devstral) and pinned Open WebUI on the existing 0.32.15 service after restart. Isolated 0.34.2 is useful but not default; official Qwen3.8 and full-reboot candidate acceptance remain open. The tested host's Ollama listener was LAN-reachable, not local-only. Released llama.cpp v0.4.1 passed bounded direct/server/HIP controls; this does not qualify every model, long-context shape or maximum-memory allocation.
Qwen3.8 27B runs locally on AMD Ryzen AI MAX+ 395 / Radeon 8060S Strix Halo systems. The useful question is no longer only “does it run?” It is which official, stock, MTP, DFlash, ROCmFP4, or performance-fork route fits the workload—and which published numbers are actually comparable.
Evidence reviewed: November 9, 2026.
Fast Answer
| Question | Current answer |
|---|---|
| Easiest measured official route | qwen3.8:27b through Ollama 0.32.13 and Vulkan/RADV |
| Guide-measured warm result | 292.49 prompt t/s and 20.42 generation t/s over nine repeats, measured as Ollama API with Ollama-default MTP drafting (draft_num_predict 4; draft-mtp logged in the same-service 64K run; per-run log for the 4K warm runs not captured); not a no-draft result; the matched no-draft control is the next row |
| Matched no-draft control (separate run) | Without MTP 12.89 generation t/s versus 22.71 t/s with Ollama-default MTP on the same model blob (Ollama 0.32.15, 2026-09-26, routine background load, nine warm repeats each); not strict-clean, and the two values are not combined with the 20.42 run. Raw evidence |
| Guide-measured capabilities | Image, tool-call, thinking, and exact retrieval through 50,059 prompt tokens passed |
| Long-context boundary | A 56,051-token attempt caused a recoverable device loss on that exact stack; separate corrected GMKtec evidence reached 261,130 evaluated tokens |
| 52-65 t/s posts | Advanced fork/quant/speculation leads that require their exact artifacts, prompts, context behavior, and independent reproduction |
For commands, the complete matrix, caveats, upstream alerts, and raw evidence, read the canonical Qwen3.8 Strix Halo decision page.
Start With Ollama
The measured official route used the 17.7GB-decimal Q4_K_M artifact:
# Requires Ollama 0.32.12 or later; measured on 0.32.13
ollama run qwen3.8:27b
The Strix Halo service still needs the guide's Vulkan/iGPU environment,
including OLLAMA_VULKAN=1 and OLLAMA_IGPU_ENABLE=1. Ollama 0.40.15 is the latest
stable release checked (November 9, 2026; availability only, not qualified here).
Ollama 0.34.2 passed isolated controls on September 19. Neither observation
inherits the measured 0.32.13 result or full normal-service/reboot qualification.
See the advisory recheck.
Why The Speed Claims Differ
Official Ollama (which enables its default MTP drafting for qwen3.8:27b), stock llama.cpp, custom quants, ROCmFP4, native MTP,
DFlash2, and adaptive speculation are different routes. Code versus prose,
cold versus cached context, context depth, generated-token count, and draft
acceptance can change the result again. A useful comparison therefore records
the entire profile, not only a tokens-per-second screenshot.
The guide keeps:
- first-party raw Ollama evidence,
- community/runtime provenance, and
- the machine-readable route matrix
separate so readers can decide which result applies to them.
What Should Be Tested Next?
The missing proof is a matched same-host ladder: stock b10687 Vulkan without speculation, native MTP, corrected HIP correctness, a published ROCmFP4 route, and a fully published DFlash/adaptive route. It must cover code and prose, 4K/16K/50K context, exact outputs, vision/tools, acceptance, memory, hashes, and raw logs.
See the live test queue or submit a benchmark report.
Independence And Affiliate Disclosure
This guide contains no affiliate links as of September 19, 2026. Future affiliate, loaned, gifted, sponsored, or early-access relationships must be disclosed near the relevant links/results and do not buy positive conclusions. Community results remain separate from first-party measurements. The public affiliate link registry remains the audit source if that status changes.