New Delhi: Alibaba-backed Qwen team has unveiled its latest open-source coding model, Qwen3-Coder-480B-A35B-Instruct, aiming to push the boundaries of agentic AI development. Built for real-world software engineering tasks and massive-scale repositories, this new large language model introduces improvements in long-context reasoning, tool use, and execution-driven learning.
Qwen3-Coder is the team’s most powerful and capable coding model so far. It targets tasks that go beyond writing code snippets, with a focus on making decisions, using developer tools, and solving software bugs over multiple steps.
Read More: Alibaba’s latest Qwen3 AI model goes big on long-text, logic and languages
Qwen3-Coder brings 480B parameters and 256K context window
Qwen3-Coder-480B-A35B-Instruct runs on a Mixture-of-Experts (MoE) setup with 480 billion total parameters, out of which 35 billion are active at any given time. This helps keep the model efficient during use while still tapping into a massive reservoir of training.
It supports 256,000 tokens as context length by default and can stretch up to 1 million tokens using extrapolation methods like YaRN. This lets it handle entire codebases, large pull requests, or documentation-heavy files with ease.
SWE-Benchmark
Beats other open models in agentic coding tasks
According to the Qwen3 blog, this version achieves state-of-the-art performance on open benchmarks for Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use. In these tests, Qwen3-Coder performs at a level comparable to Claude Sonnet 4, a closed-source high-end model.
The team says this is possible because of their work on scaling reinforcement learning, especially for tasks that are difficult to solve but easy to verify. These include automated test case creation and code fixes based on environment feedback.
Post-training uses reinforcement learning at scale
Instead of sticking to leaderboard-style coding competitions, Qwen3-Coder was trained on broader and more realistic tasks using reinforcement learning. This involved training the model on a wide variety of test cases and planning-based workflows.
In software engineering environments like SWE-Bench, the model needs to make multi-turn decisions, use tools, and correct itself based on feedback. To train it for such cases, Qwen used long-horizon reinforcement learning and deployed 20,000 environments in parallel using Alibaba Cloud.
Qwen Code CLI tool for agentic coding
Alongside the model, the team has released a command-line interface called Qwen Code. This tool is adapted from Gemini CLI but reworked to support Qwen3-Coder’s agentic reasoning abilities. It features a custom function calling setup and parser.
Users can access the Qwen3-Coder API through Alibaba Cloud’s Model Studio. The tool is meant for research use, with the hope that developers integrate the model into wider developer workflows over time.
Looking ahead: More sizes, less cost
Qwen says that smaller versions of Qwen3-Coder are in development and will be released soon. These are expected to retain strong performance while being easier to deploy and cheaper to run.
The team also teased future experiments where the model may attempt to improve itself. While that is still an open question, they are optimistic about the potential of agent-based AI in automating tough, boring parts of modern coding workflows.
According to the official model card on Hugging Face, Qwen3-Coder supports function calling, 262K context length, and includes 160 experts with 8 activated per run. It does not support thinking blocks like some other agentic models but runs without needing a flag for enabling thinking mode.
Qwen3-Coder is part of the broader Qwen3 series, which includes models trained on 7.5 trillion tokens with a 70 percent code ratio. It builds on previous work from Qwen2.5 and aims to improve both quality and alignment using cleaned synthetic data.
With this launch, the Qwen team signals it wants to compete not just in benchmarks but in real-world developer tooling, where automation and reliability matter most.
Qwen3-Coder is available on Hugging Face and Alibaba Cloud Model Studio as of July 23, 2025.