Qwen 3 235B A22B Thinking 2507

• 🧠 Reasoning-only model tuned for deep, multi-step problem solving
• 🏢 Built by Alibaba's Qwen team, Apache 2.0 licensed
• 🔧 Mixture-of-Experts: 235B total, 22B activated per token
• 📏 Model card cites native 262,144-token (256K) context
• 🆕 Split from dual-mode predecessor into dedicated thinking variant
• ⚡ Served here in FP8 for efficient deployment
• 🎯 Increased "thinking length" for highly complex tasks
• 🔧 Strong tool-calling and agentic use via Qwen-Agent

Alibaba

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Qwen 3 235B A22B Thinking 2507 is a reasoning-specialized large language model from Alibaba's Qwen team, released under the Apache 2.0 license. Architecturally it is a Mixture-of-Experts transformer with 235 billion total parameters and roughly 22 billion activated per token, using 128 experts with 8 active per token across 94 layers with grouped-query attention. It always operates in "thinking" mode, emitting reasoning traces before its final answer, and is positioned for in-depth research, technical work, and long, complex documents.

The most direct family comparison is to the original Qwen3-235B-A22B, which uniquely combined thinking and non-thinking behavior in a single switchable model. With the 2507 refresh, Qwen split that design into two dedicated checkpoints: a non-thinking Qwen 3 235B A22B Instruct 2507 and this thinking-only release. Qwen describes the 2507 line as featuring significant enhancements over the previous version, including extended 256K long-context understanding, and notes this thinking variant has an increased thinking length recommended for the hardest reasoning tasks.

Per the model card, it natively handles a 262,144-token context, and an optional configuration extends inputs toward one million tokens with sparse attention. The catalog exposes a 128K-token window in an FP8 quantization for more efficient serving.

Within the broader Qwen lineup, it sits alongside vision-language siblings such as Qwen3 VL 235B and the efficiency-focused Qwen 3 Next 80b. It retains strong tool-calling and agentic integration through the Qwen-Agent framework.