qwen3-235b-a22b-instruct-2507

Qwen3-235B-A22B-Instruct-2507 is the July 2025 refresh of Qwen's flagship Mixture-of-Experts instruct model, positioned as the non-thinking-mode sibling to the original Qwen3-235B-A22B release. It is built as a causal language model with 235 billion total parameters, 22 billion activated per inference, 234 billion non-embedding parameters, and 94 layers, paired with grouped-query attention using 64 query heads. The post-training pass emphasized instruction following, logical reasoning, mathematics, science, coding, tool usage, broader long-tail knowledge across multiple languages, and more helpful behavior on subjective and open-ended prompts, while also extending the model's native 256K context handling.

In practical terms, the model is aimed at teams that need a single large instruct model for diverse workloads: long-document analysis, multilingual Q&A and writing, code generation with tool assistance, and reasoning-heavy chat where responses should be direct rather than wrapped in chain-of-thought blocks. The FP8-quantized NIM container published on the NVIDIA NGC catalog suggests an inference path tuned for efficient deployment on NVIDIA hardware, complementing the original Qwen distribution. It fits well as a general-purpose assistant backbone for enterprise applications that value long-context comprehension and broad language coverage without switching between specialized models.