Qwen 3.6 35B-A3B is the efficiency pick of Alibaba's April 2026 generation — a sparse Mixture-of-Experts vision-language model with 35B total parameters but only 3B active per token, routed through 256 experts in a hybrid Gated DeltaNet / Gated Attention architecture. It is Ollama's default `qwen3.6` tag and the most-downloaded variant of the line.
Despite the tiny active footprint it outperforms the dense Qwen 3.5 27B on several coding benchmarks, scoring 73.4 on SWE-bench Verified and 86.0 on GPQA Diamond, and generates tokens far faster than dense models of similar weight-file size. The 24 GB Q4_K_M needs ~27 GB of memory, making 32 GB Macs and 32 GB GPUs its natural home.