Qwen3 235B A22B Instruct 2507

Qwen3-235B-A22B-Instruct-2507 is an instruction-tuned, non-thinking mode large language model developed by Alibaba’s Qwen team. It is a mixture-of-experts (MoE) model with 235 billion total parameters (22B active) and is optimized for reasoning, tool use, coding, and long-context tasks.

The model is designed for:

• Complex reasoning (e.g., AIME25, HMMT25)
• Instruction following and logic tasks
• Coding (e.g., MultiPL-E, LiveCodeBench)
• Long-context comprehension (supports up to 1M tokens)
• Multilingual knowledge and creative writing

The model supports a native context length of 262,144 tokens, and can be extended up to 1,010,000 tokens using Dual Chunk Attention and sparse attention mechanisms.

While the model supports up to 1M tokens, the recommended usable context for most tasks is up to 16,384 tokens, due to memory and latency considerations.

Yes. The model is available in FP8 quantized format, which improves inference speed and reduces memory usage.

The recommended maximum output length is 16,384 tokens, aligned with guidance from the Qwen team for generation quality and stability.

Known challenges include:

• VRAM-related issues when attempting 1M context inference without proper configuration
• Slight performance tradeoffs in long contexts with sparse attention
• Some reports of alignment inconsistencies in subjective tasks

Yes. The model supports streaming generation and agentic tool use via Qwen-Agent, which provides built-in support for function-calling and tool integration through configurable MCP files.

The model has 235 billion total parameters, with 22 billion active per token using a Mixture-of-Experts architecture.

Token pricing is split between input and output:

• $0.22 per 1M input tokens
• $0.88 per 1M output tokens

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