Qwen 3.5 397B A17B Fast

Qwen3.5 397B A17B Fast is a speed-optimized serving variant of the 397B A17B sparse Mixture-of-Experts model, tuned for lower latency while keeping the same reasoning and coding capabilities. It shares the 397 billion total parameter, 17 billion active per token design and the 262,144 token native context window, extensible toward roughly one million tokens with YaRN scaling. Reported scores match the base model, including MMLU-Pro 87.8 and SWE-bench Verified 76.4. It runs in thinking mode by default and is multimodal across text, image, and video. The faster token generation makes it a fit for interactive agentic applications that need quick responses while still supporting tool calling, structured outputs, and complex problem solving across more than 200 languages.