Qwen3.8 27B is a multimodal model with a hybrid attention design and 27.3B parameters. It is provided under the Apache 2.0 license and features a context window of 262,144 tokens. It has recorded scores of 61.7 on SWE-bench Pro and 89.2 on GPQA Diamond.
In real-world measurements on a Mac Studio M3 Ultra (256GB), the generation speed of the Q4_K_M quantized build via Ollama was approximately half that of the previous generation. Qwen3.8 achieved 14.2 tok/s, compared to 28.6 tok/s previously. However, because the number of response tokens decreased by about half, the time required per response was nearly equivalent to the previous generation.
Processing was executed via Metal by the GPU, with CPU load kept to around 6W. With Unsloth's 1-bit quantized build, prompt processing of 309 tok/s and a generation speed of 27.2 tok/s were confirmed using less than 8GB of memory. While fact-checking was reported as good, a phenomenon where the model repeatedly self-corrected when generating bash commands was also observed.
Source: Run Qwen3.8 27B locally: real numbers from my Mac Studio（HN 138pt・99コメント） (HN Search (backfill), 2026-08-29)