Qwen 3.6 lands on Ollama

batiai/qwen3.6-35b has been uploaded to Ollama. It's a model directly quantized from Alibaba's Qwen 3.6 35B-A3B BF16 weights, and it supports tools and thinking.

The tags provided are iq3, iq4, q3, q4, q6, and the guided memory requirements are as follows.

• iq3/q3: 13GB, minimum 16GB RAM
• iq4/q4: 18GB, recommended 24GB RAM
• q6: 27GB, 36GB+ RAM recommended

Benchmark figures were also presented. Compared to Qwen 3.5 35B-A3B, it's explained to have improved to SWE-bench Verified 73.4(+3.4), Terminal-Bench 2.0 51.5(+11.0), and QwenWebBench 1397(+43%).

It also emphasizes its advantage compared to Gemma 4 31B, presenting SWE-bench Verified 73.4, SWE-bench Multilingual 67.2, SWE-bench Pro 49.5, Terminal-Bench 2.0 51.5, and AIME26 92.7.

Measured performance was also disclosed. On a MacBook Pro M4 Max(128GB), it recorded iq3 45.9 t/s and iq4 46.5 t/s, and it's noted that iq3 failed on tool call JSON while iq4 passed. On a Mac mini M4(16GB), iq3 was at around 2~3 t/s level, and it's stated that iq4 is not suitable.

This model is deployed as text-only. The multimodal version requires a vision tower and mmproj.gguf, but Ollama currently provides only the text tower, choosing a configuration optimized for a single file, smaller disk/memory usage, and code/RAG/tool-call-focused workloads.