Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All
What it is
Qwen3.6-35B-A3B is a coding-focused language model optimized for 'agentic' behavior—multi-step tasks like debugging real GitHub issues, running bash commands, or fixing entire repositories. Unlike general-purpose models, it's trained to think in loops: try, fail, correct, retry. Picture a junior engineer who can read error logs, edit files, and actually iterate until tests pass.
Why it matters
If you're building AI coding agents (or considering it), this changes the cost equation. You get Claude-level performance at $0.10 per million input tokens vs. $3 for Sonnet 3.5—30x cheaper. Since it's open-source, you can self-host, fine-tune on your codebase, or audit its behavior. The gap between open and closed models just collapsed for production coding assistants.
Key details
- •35B parameters, trained on mix-quality data using ActiveReplay technique—learns more from harder examples
- •Apache 2.0 license (fully commercial-use approved), available on Hugging Face and ModelScope
- •SWE-Bench Verified: 49.1% (beats Claude Sonnet 3.5's 47.6%, GPT-4o's 38.8%)
- •Cost: $0.10/M input tokens, $0.30/M output on Alibaba Cloud—significantly undercuts closed APIs
- •Built on Qwen2.5 architecture, optimized for function calling, multi-turn coding, tool use
Worth watching
Video data provided by YouTube. Videos link to youtube.com.