Qwen-3.6-35B-A3B open‑sourced
Qwen-3.6-35B-A3B — a 35B-parameter model with a 3B-parameter active slice — was released under Apache 2.0 as a multimodal, agent-capable model positioned against much larger alternatives. Social posts note its coding and agentic capabilities and list it among free models gaining traction for practical deployments (x.com) (x.com).
Qwen released Qwen3.6-35B-A3B as open weights on April 16, giving developers an Apache 2.0-licensed model aimed at coding and agent tasks. (qwen.ai) The model uses a mixture-of-experts design, a setup that keeps many parameters on disk but activates only part of them for each token. Qwen says the model has 35 billion total parameters and 3 billion active parameters. (qwen.ai) Qwen posted the weights on Hugging Face and added an official GitHub repository this week. The Hugging Face card lists Apache 2.0 licensing and support for Transformers, vLLM, SGLang, and KTransformers. (huggingface.co) (github.com) Large language models generate text by predicting the next token, and mixture-of-experts models try to cut running costs by using only a small subset of their network on each step. Qwen says that lets this release target the economics of a smaller model while competing with larger dense models on coding benchmarks. (qwen.ai) Qwen says the model is multimodal, which means it can take text and images as input, and it supports both “thinking” and “non-thinking” modes. The model card lists a native context length of 262,144 tokens, with extension past 1 million tokens. (huggingface.co) On Qwen’s published benchmarks, the model scored 73.4 on SWE-bench Verified and 51.5 on Terminal-Bench 2.0. Qwen compared those results with Gemma 4-31B at 52.0 on SWE-bench Verified and 42.9 on Terminal-Bench 2.0, while Qwen3.5-35B-A3B scored 70.0 and 40.5. (qwen.ai) The company also says the release improves “frontend workflows” and “repository-level reasoning,” which are shorthand for editing multi-file codebases and handling longer software tasks. GitHub’s README says the update was shaped by direct community feedback and adds “thinking preservation” across conversation history. (github.com) (huggingface.co) Those claims come from Qwen’s own testing, not an independent benchmark lab, and the company notes that some evaluations used internal agent scaffolds and modified benchmark setups. The blog says SWE-bench Pro results were run on a “refined benchmark,” and several coding tests used long context windows and multi-run averages. (qwen.ai) The release lands as open-weight model makers push smaller active-parameter systems that can run cheaper in production than similarly capable dense models. Qwen’s GitHub page calls Qwen3.6 the latest addition to the family and says the team has been focusing on “stability and real-world utility” over recent months. (github.com) For developers, the immediate change is simple: a newly released Apache 2.0 model with official weights, public tooling support, and benchmark claims centered on coding agents rather than general chat. (huggingface.co) (qwen.ai)
