কিউৱেন 3 ক'ডাৰ নেক্সট

Qwen's coding model for agents and local development, optimised for repository-level understanding and multi-turn tool interaction. With only 3B parameters activated out of 80B total it reaches performance comparable to models with ten to twenty times more active parameters, which is what makes it cheap enough to run in a loop. It excels at long-horizon reasoning, complex tool use and recovery from execution failures, and it does not reason.

Alibaba's stated approach here was to scale the training signal rather than the parameter count. The model is built on an existing hybrid-attention sparse base and then trained agentically at scale: continued pretraining on code- and agent-centric data, supervised fine-tuning on high-quality agent trajectories, domain-specialised expert training in software engineering, QA and web or UX work, and finally distillation of those experts back into one deployment-ready model. The training tasks were paired with executable environments throughout, so the model learned from what actually happened when it ran something rather than from descriptions of correct code.

That recipe is aimed at three specific behaviours, and Alibaba names them: long-horizon reasoning, tool use, and recovery from execution failures. The third is the one that decides whether a coding agent is usable. A model that writes a good patch but cannot read a stack trace and try again needs a human at every step; one that can recover turns a failed run into the next attempt on its own.

The lab reports over 70% on SWE-Bench Verified with an agent scaffold, competitive results on the multilingual variant and on the harder SWE-Bench Pro, and - the more interesting claim - that its SWE-Bench Pro score keeps climbing as the agent is allowed more turns. Alibaba offers that as evidence the model holds up over long multi-turn runs rather than peaking on a single shot. With only 3 billion parameters active out of 80 billion, it says the model matches or exceeds several much larger open models on agent-centric evaluations, which is what puts it on a useful efficiency frontier rather than merely at a low price.

The economics are the practical reason to reach for it: at this rate, letting the agent retry is not a budget decision. Repository-level edits, running tests, reading the failure and going round again is exactly the loop it was trained in, and the lab demonstrates it inside terminal agents, editor extensions and browser agents rather than in chat.

Be clear about what it is not. It does not reason - there is no thinking pass and no effort dial - and it is a coding specialist, not a general assistant: on general knowledge and reasoning evaluations it lands far below any of the lab's flagship models. Alibaba's own summary says there is still much room for improvement and names tool use, hard problems and complex task management as the things it wants to fix next. Use it for repetitive, verifiable code work where a test can say whether it succeeded; take a reasoning model for architecture decisions, novel algorithms, or anything that is not code.