Alibaba Ships Qwen3.6-35B-A3B Free: 3B Active Params Score 73.4% SWE-Bench, Beat Gemma 4 by 21 Points
Alibaba’s Qwen team released Qwen3.6-35B-A3B on April 16, making the weights available for free download on Hugging Face and ModelScope under Apache 2.0. The model has 35 billion total parameters but activates only 3 billion per forward pass — a sparse mixture-of-experts architecture that lets it run on 16GB of RAM without quantization, or on a standard Mac Mini with the Q4 Ollama build at roughly 23GB.
The headline number is SWE-bench Verified at 73.4%. For context, Google’s directly comparable open-weight MoE — Gemma 4 26B A4B, which activates 4 billion parameters per token — scores 52.0% on the same benchmark. Qwen3.6-35B-A3B uses less compute per token and wins by 21 points. That gap is significant: frontier closed models Claude Opus 4.6 and GPT-5.4 sit at 80.8% and 77.2% respectively, meaning the new Qwen open weight lands within 8 percentage points of OpenAI’s flagship.
Benchmark Profile
| Benchmark | Qwen3.6-35B-A3B | Gemma 4 26B A4B |
|---|---|---|
| SWE-bench Verified | 73.4% | 52.0% |
| Terminal-Bench 2.0 | competitive (Alibaba: “clearly superior” to A4B) | — |
| RefCOCO (spatial vision) | 92.0 | — |
| ODInW13 | 50.8 | — |
| Vision vs Claude Sonnet 4.5 | on par / slightly ahead on some tasks | — |
The model also supports multimodal input natively. On standard vision-language benchmarks (MMBench, RealWorldQA, SimpleVQA), Alibaba reports performance matching Claude Sonnet 4.5 — notable for a model that runs locally at no cost. Spatial intelligence scores (RefCOCO 92.0, ODInW13 50.8) place it at the front of the open-weights vision field for its active-parameter tier.
Architecture
The MoE routing means most of the 35B parameters are unused for any given token. Qwen3.6-35B-A3B implements a think/no-think toggle: in agentic tasks, leaving chain-of-thought reasoning on (via the preserve_thinking flag) materially improves performance on long-horizon agent loops; for latency-sensitive inference, turning it off is recommended. Both modes are available through the standard API.
The model loads at roughly 65GB full precision on VRAM (tested on an H100) or 23-24GB in Ollama’s Q4_K_M quantization. Independent tests show the quantized version is close but not identical to full precision on edge cases — completeness in code generation drops slightly, particularly on game logic boundary conditions. For production agentic workloads where correctness matters, full precision on appropriate hardware is the better path.
Access and Pricing
Weights are free. The API route on Alibaba Cloud Bailian carries the name qwen3.6-flash. Qwen Studio offers free interactive access. No rate-limit details have been published for the free API tier.
The Qwen3.6 family positions as a price-performance challenge to GPT-5.4 and Claude Opus 4.6 for agentic coding. Qwen3.6-Plus, the closed API variant released April 7, is priced at roughly $0.276/M input tokens — approximately 18x cheaper than Claude Opus 4.6 for long-context workloads. The open-weight 35B-A3B extends that cost advantage to zero-API-cost local deployments.
Alibaba says smaller models in the Qwen3.6 series will follow, with the team’s focus shifting to long-horizon repository-level agentic tasks.