Want smart insight into your inbox? Enterprise AI, only what matters to data and security leaders, sign up for our weekly newspapers. Subscribe now
It has been re-rejected by Chinese e-commerce giant Alibaba’s “Qwen Team”.
Free and a few days after release with open source licensing Now what is the top of non-rationing large large language model (LLM) in the world Compared to the complete break-up, even well-funded American laboratories such as Google and Openai AI models-QWEN3 -235B-A22B-2507, for a long time named, this group of AI researchers, this group has so far surfaced with another blockbuster model.
he is Qwen3-Coder-480B-A35B-Instruct, New one Open-SOLM LLM focused on assisting software development. It is designed to handle complex, multi-step coding workflow and can create complete, functional applications in it. Second Or minute.
The model is deployed to compete with proprietary offerings such as Cloud Sonnet -4 in agentic coding tasks and sets new benchmark score among open models.
It is available Throat face, Github, Qwen chatThrough Alibaba’s Quven APIAnd a growing list of third-party vibe coding and AI tool platforms.
Open sourcing licensing means low cost and high optionalism for enterprises
But unlike Cloud and other proprietary models, Qwen3-Coder, which we would call it for short, is now available. Open Source Apache 2.0 License,
It is very highly performing on the third-party benchmark and anecdote use among AI power users for “vibi coding”-using the nature language and without formal development procedures and steps-at least one, at least one, LLM researcher Sebastian RashkaIt is written on x that: “It can be the best coding model yet. The general-purpose is quiet, but if you want the best in coding, you win the expertise. No free lunch.”
Developers and enterprises interested in downloading this can be found on AI code sharing repository Throat face,
The enterprises do not have the ability to host the model on their own or on their own through various third-party cloud invention providers through cloud invention providers. Alibaba Badal through Qwen APIWhere the cost of up to 32,000 tokens per million tokens starts at $ 1/$ 5 million per million tokens (MTOK), then up to $ 1.8/$ 9128,000, $ 3/$ 15 to 256,000 by $ 3/$ 15 and $ 6/$ 60 for full million.
Model architecture and capacity
According to the document released by QWEN team online, Qwen3-Coder is a mixture-off-specialist (Moe) model with 480 billion parameters, 35 billion active per querry, and 8 out of 160.
It supports 256k token reference length, which is originally with extras up to 1 million tokens using yarn (yet another rope extractation – a technique is used to expand the reference length of a language model, which is used during revision of the rotary positive embeding (rope).
One reason designed as a language model, it has 62 layers, 96 meditation for queries and 8 for the key-value pairs. It is adapted to the following tasks and supports token-skilled, instructions and support
High performance
Qwen3-Coder has achieved leading performance among open models on several agentic assessment suits:
- SWE-Bench verified: 67.0% (standard), 69.6% (500-turn)
- GPT-4.1: 54.6%
- Gemini 2.5 Pro preview: 49.0%
- Cloud Sonnet -4: 70.4%
The model has a competitive score in tasks such as agentic browser use, multi-language programming and tool using. Visual benchmark codes show progressive reforms in recurrents in categories such as code generation, SQL programming, code editing and instruction.
Along with the model, Quven has an Open-Sound Qwen Code, a CLI tool from the Gemini code. This interface function supports calling and structured prompting, making it easier to integrate the Qwen3-Coder into coding workflows. Qwen code supports node.JS environment and can be installed from NPM or source.
Qwen3-Coder also integrates with developer platforms such as:
- Cloud Code
- Cline (as an openi-composed backnd)
- OLLEMA, LMSTUDIO, Mlx-LM, LLAMA.CPP, and Ktransformers
Developers can run Qwen3-Coder locally or connect through Openai-Sangat API using the endpoints hosted on Alibaba Cloud.
Post-training technology: Code RL and long-hurizon planning
In addition to pretering on 7.5 trillion tokens (70% code), Qwen3-Coder Benefits from advanced post-training techniques:
- Code RL (Learning reinforcement): Miscellaneous, verification of verification emphasizes high quality, execution-making learning on tasks
- Long-Horizone agent RL: Trains the model to plan, use tools and optimize multi-turn interactions
This phase simulates real -world software engineering challenges. To enable this, Qwen created a 20,000-environmental system on Alibaba Cloud, offering the required scale for the evaluation and training model of the model on the complex workflows found in Swe-Bench.
Enterprises implication: AI for Engineering and Devops Workflow
For enterprises, Qwen3-Coder offers an open, highly capable option for closed-source ownership models. With strong results in coding execution and long reference logic, it is particularly relevant:
- Codebase-level understanding: Ideal for AI system who should understand large repository, technical documentation or architectural patterns
- Automatic Bridge Request Workflow: Its ability to plan and adapt to the turn makes it suitable for reviewing auto-generating or bridge requests
- Equipment Integration and Orciliation: Through its original tool-coaling API and function interfaces, the model can be embedded in internal tooling and CI/CD systems. This makes agentic workflows and feasible for products, that is, they triggers one or several tasks where the user triggers one or several tasks that the AI model is closed and autonomally, on its own, when finished or when the question arises.
- Data residence and cost control: As an open model, the enterprise can deploy Qwen3-Coder on its own infrastructure-Claude-country or on-ride-vynder lock-in and use more directly
Support for long references in various deity environment and supports for modular regional options, QWEN3-Coder makes a candidate for both large technical companies and small engineering teams production-grade AI pipelines.
Developer access and best practices
To use Qwen3-Coder, QWEN recommends:
- Sampling settings: temperature = 0.7, top_p = 0.8, top_k = 20, repetition_penalty = 1.05
- Output length: up to 65,536 tokens
- Transformer version: 4.51.0 or later (older versions can throw errors due to qwen3_moe inconsistency)
API and SDK examples are provided using Openai-compatible python customers.
Developers can define custom tools and invite Qwen3-Coder dynamically during interaction or code generation functions.
AI power users war
Initial reactions for Qwen3-Coder-480B-A35B-Instruct have been particularly positive among AI researchers, engineers and developers who have tested the model in the real world coding workflow.
In addition to the high appreciation of Raschka above, Volfram Revenwolf, an AI engineer and evaluator in Almindai, shared their experience Equipment of model with cloud code on Xby stating, “This is definitely the best in the present.”
After testing several integration proxy, Revenwolf stated that he eventually produced his own using Littlem to ensure optimal performance, which demonstrates the model’s appeal to doctors focusing on toolchen adaptation.
Teacher and AI Tinker Kevin Nelson also weighed at X After using the model for simulation works.
“Qwen 3 Kodar is at another level,” He posted, given that the model not only executed on the scaffolds provided, but also embedded a message within the output of simulation – an unexpected but welcome indication of the awareness of the model of the work reference.
Even Twitter co-founder and founder of Square (now called “Block”), Jack Dorsi posted an X message in praise of the model, Write,Goose + Qwen3-Coder = Wow,“In terms of its block’s open source AI agent Framework Goose, which was covered back in January 2025.
These reactions suggest that the Qwen3-Coder is resonating with a technically lover user base, seeking performance, adaptability and intensive integration with the existing growth stack.
Looking forward: more size, more use cases
While this release focuses on the most powerful version, Qwen3-Coder-480B-A35B-Instruct, Qwen team indicates that additional models are in size development.
These will aim to comprehend access, with the cost of low deployment, to offer equal capabilities.
Future work also involves the discovery of self-reform, as the team examines whether agentic models can refine their own performance through the use of the real world.
