News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnQwen3-235B-A22B

Qwen3-235B-A22B

Mixture-of-experts model with switchable reasoning modes

Qwen3-235B-A22B is a large language model released by the Qwen team in April 2025. It uses a mixture of experts architecture to balance efficiency and power, allowing users to toggle between deep reasoning and fast responses within a single system.

Share on WhatsApp

  • Total 235B parameters with 22B activated per token
  • Supports 119 languages and dialects
  • Switchable thinking and non-thinking modes
  • Context length up to 131,072 tokens with YaRN
  • Released under Apache 2.0 licence

What it is

Qwen3-235B-A22B is a large language model from the Qwen family, announced in April 2025. It features a total of 235 billion parameters but activates only 22 billion per token through a mixture of experts design. This structure allows it to handle complex logical tasks while remaining efficient for general dialogue. The model supports over 100 languages and integrates tools for agent-based workflows.

How it came about

The Qwen Team released this model as part of the third generation of their series. It follows Qwen2.5 and introduces a unified framework that combines reasoning and non-reasoning capabilities. Earlier versions required switching between different models for chat or logic tasks, but this iteration merges them. The technical report was published in May 2025, detailing its advancements in multilingual support and efficiency.

Inside the model

The model uses a mixture of experts architecture with eight activated experts per token. It natively supports a context window of 32,768 tokens, extendable to 131,072 using YaRN scaling. A key feature is the ability to switch between thinking mode for complex problems and non-thinking mode for speed. Users can control this via API settings or by adding specific tags like /think in prompts. Training included post-training alignment to improve instruction following.

What it can do

It performs strongly in mathematics, code generation, and commonsense reasoning when in thinking mode. In non-thinking mode, it matches previous instruct models for general conversation. The system excels in agent tasks, allowing precise integration with external tools. Evaluations show competitive results against other top-tier models in coding and math benchmarks. It supports 119 languages and dialects, improving cross-lingual understanding.

How to use it

Weights are publicly available under the Apache 2.0 licence. Developers can run it locally using frameworks like Ollama, llama.cpp, or vLLM. The model supports OpenAI-compatible API endpoints through SGLang or vLLM servers. Documentation provides guidance on hardware requirements and inference performance for deployment.

Questions people ask about Qwen3-235B-A22B

What is the difference between thinking and non-thinking mode?

Thinking mode engages deeper reasoning for complex logic, math, and coding tasks. Non-thinking mode provides faster, efficient responses for general dialogue. Users can switch between them using API settings or prompt tags.

How many parameters does the model use during inference?

While the model has 235 billion total parameters, it activates only 22 billion per token. This mixture of experts approach reduces computational cost while maintaining high performance.

Can I use this model for commercial purposes?

Yes, the model weights are released under the Apache 2.0 licence. This permissive licence allows for commercial use, modification, and distribution without significant restrictions.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Qwen3-235B-A22B yet, so the facts on this page come from the sources listed above: huggingface.co/Qwen/Qwen3-235B-A22B, arxiv.org/abs/2505.09388, qwenlm.github.io/blog/qwen3/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.