Qwen3 235B A22B Thinking 2507

Qwen3-235B-A22B-Thinking-2507 is an open-weight large language model (LLM) developed by the Qwen team. It is a reasoning-optimized variant of the Qwen3-235B-A22B MoE (Mixture of Experts) model, released in July 2025, with significant enhancements in long-context reasoning, tool usage, and alignment capabilities.

This model is designed for:

• Complex reasoning tasks (math, logic, science, coding)
• Academic and professional benchmarks
• Agentic applications requiring tool use
• Long-context use cases, including document analysis and chain-of-thought workflows
• Creative writing and instruction following.

The native maximum context length is 262,144 tokens. With custom configuration and memory requirements, it can support up to 1 million tokens using Dual Chunk Attention and sparse attention techniques.

Usable context depends on deployment setup. For most use cases, 262K tokens are supported natively. For ultra-long context (approaching 1M tokens), you must reconfigure the model and allocate ≥1000 GB of GPU memory.

The model performs best with a temperature of 0.6, TopP=0.95, and TopK=20.

The recommended output length is:

• 32,768 tokens for typical queries
• Up to 81,920 tokens for high-complexity tasks (e.g., math/code reasoning).

Known issues include:

• Memory constraints when using long contexts (OOM errors)
• Performance degradation if best practices for prompt and output format aren't followed

Yes. The model supports streaming generation and tool use via Qwen-Agent, which handles function-calling templates and tool parsers.

• Total parameters: 235B
• Activated parameters: 22B per forward pass (MoE)
• Experts: 128 total, 8 active per token.

The model is released under the Apache 2.0 license, permitting commercial use with attribution.