AI BriefingQwen3.6·gpt-oss Benchmark
Key point
On MBP M3 Max benchmarks, the gap between three Qwen3.6 variants and gpt-oss:120b was revealed.
Details
On an MBP M3 Max with Ollama, three Qwen3.6 variants and gpt-oss:120b were compared using temp 0.2, --think=false, and a structured research-brief workload.
qwen3.6:35b-a3b-coding-nvfp4: 6 seconds, 21GBqwen3.6:35b-a3b-q8_0 (MoE): 22 seconds, 38GBqwen3.6:27b-q8_0 (Dense): 67 seconds, 29GBgpt-oss:120b: 61 seconds, 65GB
--think=false was correctly applied to all three Qwen3.6 models, but was ignored by gpt-oss, causing reasoning traces to be printed directly to stdout.
A Modelfile overlay that only changed FROM model and PARAMETER temperature 0.2 reused the existing layers, adding almost no extra disk usage. On the same workload, the 35B-A3B MoE was about 3x faster than the 27B dense model.
The text-only coding-NVFP4 model doesn't reject images fed in via the API and can instead generate false descriptions, so models allowed to receive images: should be restricted to a separate allowlist.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NVIDIA Open-Sources Local AI Inference Router 'PAIR'
Why LLM Research Agents Generalize Without Benchmark Overfitting: 'Compressibility' Is Key
Cerebras Model Catalog (Website)