Qwen3 235b A22B Instruct 2507

Qwen3 235B A22B Instruct 2507 is the non-thinking, instruction-tuned member of Alibaba's Qwen3 family, built as a Mixture-of-Experts model with 235B total parameters and 22B activated per pass. It was developed alongside a separate "thinking" variant in response to developer feedback, aiming to deliver stronger general instruction following and reasoning without requiring extended chain-of-thought prompting. The Together AI listing describes it as an enhanced Qwen3 release optimized for cost-efficient high-throughput inference with FP8 weights, positioning it as a practical workhorse for production chat workloads rather than a deep-reasoning specialist.

The model stands out for its unusually long context handling and competitive benchmark profile among non-reasoning systems. Together AI documents a 262K-token context window suitable for large document analysis, long-running agent traces, and retrieval-augmented workflows, while the Cerebras deployment serves a 131K context with FP8 weights from US data centers. On the Artificial Analysis Intelligence Index, a blended score across general knowledge, reasoning, coding, and STEM benchmarks, Qwen3 235B 2507 Instruct is reported to outperform GPT-4.1, Claude Opus 4, DeepSeek V3, and Kimi K2, reaching state-of-the-art results among non-reasoning models. Combined with tool calling and temperature control, it fits teams that need a fast, long-context chat model for assistants, code helpers, and agent pipelines where deep deliberation is unnecessary but breadth and reliability matter.