Try it Out!
Try it in the AI Studio playground (instruct model) Try with API: qwen3_2507_1.ipynbTL;DR
- Qwen3-235B-A22B-2507 (released July 2025) is the updated version of Qwen3-235B-A22B (April 2025).
- Two separate model variants are now available:
- This marks a shift from the hybrid “Thinking” mode used in previous releases.
-
From the Qwen team (source):
“After talking with the community and thinking it through, we decided to stop using hybrid thinking mode. Instead, we’ll train Instruct and Thinking models separately so we can get the best quality possible.” “A small update to Qwen3-235B-A22B, but a big improvement in quality!”
Highlights:
- Released: July 2025
- 235B total parameters / 22B active (Mixture of Experts architecture)
- Native 262K token context window
- Strong performance on math, logic, and coding tasks
- Outperforms Kimi K2 (1T params) and Claude Opus 4 (Non-Thinking) in several key benchmarks
- Try it on Nebius AI Studio
- License: Apache 2.0
Fun Fact
Qwen3-235B-A22B-2507 uses a Mixture of Experts (MoE) architecture with 235B total parameters — but only 22B are active at any given time! So you get massive model performance with much lower compute costs. Smart and efficient.Performance and Benchmarks
Official benchmarks
From the Qwen team:
Artificial Analysis Benchmark
According to Artificial Analysis, Qwen3 is currently the top-performing open-source model — and it’s getting close to matching proprietary SOTA models.
Fun Benchmark: “Pelican riding a bicycle”
Inspired by Simon Willison’s fun experiment (see here), this benchmark is all about how well models generate quirky, imaginative responses. Prompt:Generate an SVG of a pelican riding a bicycleYou can see our full pelican tests here. So how does gpt-oss do? Let’s see.
| Old Version: Qwen3-235B-A22B (April 2025) | New Version: Qwen3-235B-A22B-Instruct-2507 (July 2025) |
|---|---|
 |  |