Do you want smarter insights in your inbox? Sign up for our weekly newsletters to get what is essential for AI leaders, data and security. Subscribe now
“Qwen team” of the Chinese giant e-commerce Alibaba did it again.
Just days after release for free and licensing Open Source Which is now best an unjustified model of a large language (LLM) in the world -a full stop, even compared to the reserved AI models from well-financed American laboratories, such as Google and OpenAi-in the form of a long name called Qwen3-235B-A22B-2507, this group of AI researchers came out with another hit model.
This is QWEN3-Koder-480B-A35B-InstructIN novel Open-source LLM focused on help in creating software. It has been designed to support sophisticated, multi -stage work flows and can create full -fledged, functional applications in seconds or minutes.
The model is set to compete with reserved offers, such as Claude Sonnet-4 in agency coding tasks and sets novel comparative results among open models.
Is available HuggingIN GirubIN Chat QwenBy Alibaba’s qwen APIand a growing list of atmospheric coding and AI tool platforms.
Open licenses for acquisition mean inexpensive and high optional for enterprises
But unlike Claude and other reserved models, Qwen3-Coder, which we will call it, is now available at Open Source Apache 2.0 license licenseWhich means that every enterprise can treat without fees, download, modify, implement and utilize in their commercial applications for employees or end clients without paying Alibab or anyone else.
It is also so productive on third-party references and anecdotal utilize among advanced AI users to “coding the atmosphere”-dialing using natural language and without formal processes and development steps-at least one, at least one, LLLM researcher Sebastian RakkaHe wrote on X that: “This can be the best coding model so far. The overall way is cool, but if you want the best in coding, the specialization wins. No free lunch.”
Developers and enterprises interested in its download can find the code in the AI code repository Hugging.
Enterprises that do not want or do not have the ability to host the model by themselves or through various suppliers of inference in the cloud of other companies, can also utilize it directly by Alibaba Cloud Qwen APIWhen the costs of tokens per million start from USD 1/5 for one million tokens (MTOK) for entry/output of up to USD 32,000, followed by $ 1.8/9 for USD 128,000, 3/15 USD for up to USD 256,000 and 6 USD/60 USD for a full million.
Architecture model and possibilities
According to the documentation published by the QWEN Online team, Qwen3-Coder is a model of expert mixture with 480 billion parameters, 35 billion energetic on request and 8 energetic experts at 160.
It natively supports the length of the context of 256K tokens, with extrapolation up to 1 million tokens with yarn (another lines extrapolation – the technique used to expand the length of the language model outside its original training limit by modifying rotating embedded positions (ROPE) used during calculations of attention. This capacity exhausts the model for understanding and manipulating the entire repositories or the length of documents or the length of documents in one pass.
Designed as a causal language model, it has 62 layers, 96 heads of attention to queries and 8 for key couples. It is optimized for savings based on the instructions and skipped service
High performance
The QWEN3 coder has achieved leading performance among open models in several agency rating packages:
- Its verified: 67.0% (standard), 69.6% (500 turns)
- GPT-4.1: 54.6%
- Gemini 2.5 Pro Preview: 49.0%
- Claude Sonnet-4: 70.4%
The model also assesses competitively between tasks such as the utilize of an agency browser, multilingual programming and the utilize of tools. Visual reference points show progressive improvement in training iterations in terms of such as code generation, SQL programming, code editing and observation of instructions.
In addition to the Qwen model, it has Open-Sourced Qwen Code, a tool of the CLI tool for the Gemini code. This interface supports function call and structural hint, facilitating the integration of the QWEN3 coder code with coding flows. The QWEN code supports node.js and can be installed using NPM or from the source.
Qwen3-Koder also integrates programmers’ platforms such as:
- Claude code (by adjusting the proxy or adapting the village of view)
- CLINE (as a backend compatible with OpenAI)
- Ollam, Lmstudio, MLX-LM, Lamam.CPP and KTRansformers
Developers can launch QWEN3-Koder locally or connect via API interfaces compatible with OPENAI using end points hosted in Alibaba Cloud.
Techniques after training: RL code and long -term planning
In addition to pre -claim at 7.5 trillion tokens (code 70%), QWEN3 Coder uses advanced techniques after training:
- Code RL (reinforcement learning): emphasizes high -quality learning to diverse, verifiable code tasks
- Long Horizon Agent RL: Training the model for planning, using tools and adaptation of interaction with many ends
This phase simulates the real challenges of software engineering. To turn it on, Qwen has built a system of 20,000 environments on Alibaba Cloud, offering the scale necessary for the assessment and training of models on sophisticated work flows, such as those found in SW.
Enterprise implications: AI for the flows of engineering and devops
In the case of qwen3-koder companies, it offers an open, highly talented alternative to reserved closed models. Thanks to the powerful results in coding and the reasoning of the long -time, this is particularly essential for:
- Code understanding: Ideal for AI systems that must understand gigantic repositories, technical documentation or architectural patterns
- Automatized work flows Pull demands: His ability to plan and adapt in corners makes it suitable for automatic generation or review of pulling demands
- Integration and orchestration of tools: Thanks to the native API interface and functions, the model can be embedded in internal tool and CD tool and systems. This makes it particularly profitable in the case of agency work flows and products, i.e. those in which the user launches one or many tasks that he wants the AI model to leave and perform autonomously independently, checking only after completing or when questions arise.
- Data residence and cost control: As an open model, enterprises can implement QWEN3-KODER on their own infrastructure-is it native in the cloud or location-suppliers of suppliers and computing utilize of computing utilize more directly direct
The service of long contexts and modular implementation options in various DeV environments means that QWEN3-Koder is a candidate for the production of AI pipelines in both gigantic technology companies and smaller engineering teams.
Access to programmers and the best practices
To optimally utilize QWEN3-Koder, QWEN recommends:
- Sampling settings: temperature = 0.7, TOP_P = 0.8, TOP_K = 20, repetition_Penalty = 1.05
- Exit length: up to 65,536 tokens
- Transformers version: 4.51.0 or newer (older versions may throw errors due to the incompatibility of Qwen3_MOE)
Examples of API and SDK are delivered using Python clients compatible with OpenAI.
Developers can define non-standard tools and allow QWEN3-KODER to recall them dynamically during code generation tasks.
A heated early party from AI Energy users
The initial answers to instructor’s instructor instructorts QWEN3-Coder-480B-A35B were clearly positive among AI researchers, engineers and programmers who tested the model in real coding flows.
In addition to the lofty praise of Raschka above, Wolfram Ravenwolf, AI engineer and evaluator in ELLAMindai, shared his experience Model integration with Claude code on xstating “It is certainly the best now.”
After testing several integration proxies, Ravenwolf said that he finally built his own using Liteellm to ensure optimal performance, showing the model’s appeal to practical practitioners focusing on the adaptation of the tools.
Educator and AI Tinkerer Kevin Nelson also weighed x After using the model for simulation tasks.
“Coder Qwen 3 is at a different level” He wrote, noting that the model made not only on the given scaffoldings, but even set a message as a result of simulation – an unexpected but welcome sign of the model’s consciousness about the context of tasks.
Even co -founder and square (currently called the “block”), Jack Dorse, published a message X for the praise of the model, writing: “Goose + qwen3-coder = wow,“In relation to Open Source AI AGWORK Framework Framework, which Venturebeat discussed in January 2025.
These answers suggest that the QWEN3-Koder resonates with a technically experienced user base, looking for performance, adaptive ability and deeper integration with existing programming piles.
Looking to the future: more sizes, more cases of utilize
While this version focuses on the most powerful variant, Qwen3-Koder-480B-A35B-Instruct, the QWEN syndrome indicates that the additional size of the models is in development.
They will strive to offer similar possibilities with lower implementation costs, expanding availability.
Future work also includes a test of self -improvement, because the team is investigating whether agency models can easily improve their own results by using in the real world.
