Qwen3 235B A22B Thinking 2507 (NovitaAI)

Qwen3-235B-A22B-Thinking-2507 is a thinking-mode causal language model from the Qwen3 family, built on a Mixture-of-Experts architecture with 235 billion total parameters and 22 billion activated per pass. It spans 94 layers with 128 experts of which 8 fire per token, and its non-embedding parameter count reaches 234 billion. Training covers both pretraining and post-training stages, and the model is purpose-built for extended chain-of-thought reasoning rather than quick conversational replies, running only in thinking mode by default.

The 2507 release sharpens reasoning quality on tasks that typically require expert-level thinking, including logic, mathematics, science, coding, and academic benchmarks, while also improving general capabilities such as instruction following, tool usage, text generation, and human preference alignment. It natively handles a 262,144-token context, giving it strong long-document understanding, and it has been reported to reach 92.3% on AIME25 alongside claims of outperforming several leading closed reasoning models. The model fits workloads that need deep, deliberate problem solving, such as complex math proofs, code analysis, and multi-step research over very long inputs, and it is available through the Novita AI platform alongside the original open-source release.