The “Qwen team” by Chinese e-commerce giant Alibaba did it again.
Only days after the free and with open source license What is the highest that does the big voice model (LLM) that has not been implemented on this planet -Top, even in comparison with proprietary AI laboratories similar to Google and Openai-in type of the long-named QWEN3-235B-A22B-25507, this group of AI researchers has published one other blockbuster model.
That means QWen3-cooder-480b-A35BPresent a brand new one Open source LLM focused on the support in software development. It is designed for complex, multi-stage coding workflows and may create full-fledged, functional applications in or in minutes.
The model is positioned in such a way that you simply compete with proprietary offers similar to Claude Sonnet-4 in agent coding tasks and set recent benchmark reviews between open models.
It's available on HugPresent GirubPresent Qwen chatabove Alibabas Qwen APIAnd a growing list of vibe coding and AI tool platforms from third-party providers.
Open sourcing licensing means inexpensive and high options for firms
But in contrast to Claude and other proprietary models, QWen3 code, which it calls it, is now available under one Open Source Apache 2.0 licenseThis implies that it is freed from charge for each company to take over, change, change and use of their business applications for workers or end customers without paying alibaba or other others.
It can also be so high in benchmarks of third-party providers and the anecdotal use in AI power users for “vibe coding” coding with natural language and without formal development processes and steps, a minimum of one. LLLM researcher Sebastian Rakkawrote that about X:
Developers and corporations which can be keen on downloading him Hug.
Companies that are not looking for to record the model alone or through various cloud inferior providers of third-party providers may use it directly Through the Alibaba Cloud Qwen APIWhere the fee per million is at 1/5 USD per million tokens (MTOK) for input/output of as much as 32,000 tokens, then $ 1.8/9 US dollars for as much as $ 128,000, $ 3/15 for as much as $ 256,000 and $ 6/60.
Model architecture and skills
According to the documentary published by QWen Team online, QWen3-cover is a mix mixture mixture (Mie-of-experts) with 480 billion total parameters, 35 billion lively per query and eight lively experts of 160.
It supports 256 -kk -token context native, whereby the extrapolation as much as 1 million tokens using yarn (one other rope rextrapolation using the context length of a voice model expands the rotation positioning (rope) used in the course of the attention calculation.
It was developed as a causal voice model and has 62 levels, 96 attention heads for queries and eight for key value pairs. It is optimized for token efficient
High performance
QWen3 code has achieved a number one performance in open models in several agent assessment suites:
- SWE-bench verified: 67.0% (standard), 69.6% (500 turn)
- GPT-4.1: 54.6%
- Gemini 2.5 per preview: 49.0%
- Claude Sonnet-4: 70.4%
The model also evaluates competitive via tasks similar to agent browser use, multilingual programming and using tools. Visual benchmarks show a progressive improvement about training literations in categories similar to codegen, SQL programming, code processing and directions.
In addition to the model, Qwen has the open sourcing QWen code, a Cli tool that was made out of Gemini code. This interface supports the function call and structured request, in order that QWen3 code in coding workflows into the mixing of QWen3 cooders becomes easier. QWen code supports Node.js environments and will be installed via NPM or from source.
QWen3 code also integrates into developer platforms similar to:
- Claude code (via Dashscope -Proxy or Router adjustment)
- CLINE (as an Openai-compatible backend)
- Ollama, LMStudio, MLX-LM, Lama.cpp and Ktransformers
Developers can do QWen3 coders locally or establish a connection via Openai-compatible APIs using endpoints which can be hosted on Alibaba cloud.
Post-training techniques: code RL and long-term planning
In addition to preparing for 70% code (70% code), QWen3 coders profit from advanced night training techniques:
- Code RL (learning reinforcement): emphasizes high -quality, execution -driven learning with various, verifiable code tasks
- Long-Horizon agent RL: train the model to plan, use and adapt tools and adapt to multiturn interactions
This phase simulates real software engineering challenges. To activate this, QWen built a 20,000-environment system in Alibaba Cloud and offered the dimensions that’s required for the evaluation and training of models for complex workflows as in SWE-Bench.
Implications for firms: AI for engineering and devops workflows
For firms, QWen3 code offers an open, top-class alternative to closed proprietary models. With strong results when coding the execution and the long context pondering, it is especially relevant for:
- Understanding on the code base level: Ideal for AI systems which have to know large repositories, technical documentation or architectural patterns
- Automated pull request -Workflows: The ability to plan and adapt to the rounds
- Tool integration and orchestration: Thanks to its native tool API and functional interface, the model will be embedded in internal tool and CI/CD systems. This makes it particularly practical for agents workflows and products, ie those through which the user triggers a number of tasks that triggers the AI model and does it itself, and only checks when it’s over or when questions arise.
- Data residence and price control: As an open model, firms can provide QWen3 code for their very own infrastructure.
The support for long contexts and modular provision options in various development environments makes QWen3 coders a candidate for AI pipelines for production qualities in large tech firms and smaller engineering teams.
Developer access and best practices
To optimally use QWen3 code, Qwen recommends:
- STONTINE POSSION: Temperature = 0.7, TOP_P = 0.8, TOP_K = 20, repetition_penalty = 1.05
- Exit length: as much as 65,536 tokens
- Transformer version: 4.51.0 or higher (older versions can throw errors attributable to QWEN3_Moe -inkomatibility)
APIS and SDK examples are provided using Openai-compatible Python clients.
Developers can define user-defined tools and have QWen3 coders dynamically accessed in the course of the conversation or code of code.
Warm early reception of KI -Power users
Initial answers to QWen3-cooder-480B-A35B-Instruct were significantly positive amongst AI researchers, engineers and developers who tested the model in real coding workflows.
Wolfram Ravenwolf, a AI engineer and viewer at Ellamindai, shared his experience with Raschkas Lofty above, but in addition his experience Integration of the model in Claude Code on XIndication,
After Ravenwolf tested several integration poxies, he finally said with Litellm his own to make sure optimal performance, and the model of the model to practical practitioners who concentrate on the toolchain adaptation.
Pedagogue and AI Tinkerer Kevin Nelson also has xgas X After using the model for simulation tasks.
He posted and located that the model was not only executed on the scaffolding, but even a message embedded within the edition of the simulation – an unexpected but welcome sign for the notice of the model for the duty context.
Even Twitter co-founder and Square (now known as a “block”), Jack Dorsey published an X message in regards to the praise of the model. Write: “” With regard to the open source Ai Agent Framework Gans of his block, which Venturebeat covered in January 2025.
These answers indicate that QWEN3 code with a technically versed user-based performance, adaptability and deeper integration with existing development stacks is resonance.
View the foresight: more sizes, more applications
While this publication focuses on essentially the most powerful variant, QWen3-cooder-480b-A35B instruct, the QWen team indicates that there are additional model sizes in development.
These are intended to supply similar skills with lower provision costs and expand accessibility.
Future work also includes research into self -improvement, for the reason that team examines whether agent models can refine its own performance through real use.
