Qwen3.5 397B-A17B

Qwen3.5 397B A17B by Alibaba Qwen is a multimodal model released in February 2026 with a 262K token context window (approximately 131 books) and an open-source license. The Qwen3.5 series 397B-A17B native vision-language model is built on a hybrid architecture integrating a linear attention mechanism with a sparse mixt Top benchmark scores include Chatbot Arena Elo Overall at 1443.7, Chatbot Arena Elo Coding at 1394.5, and OpenCompass AIME2025 at 92.3%. Additional results include MMLU-Pro at 87.6%, GPQA-Diamond at 88.4%, LiveCodeBenchV6 at 83.0, and Humanity's Last Exam at 27.5%. The model demonstrates particularly strong performance

Qwen3.5 is a flagship large language model series launched by Alibaba on February 16, 2026, comprising two versions: Qwen3.5-Plus and Qwen3.5-397B-A17B. The series supports text and multimodal tasks, utilizing a hybrid attention mechanism and a sparse mixture-of-experts (MoE) architecture, and is optimized for logical Alibaba announced the Qwen3.5 development plan in January 2026, with its code merge request appearing on Hugging Face on February 9 and showcases at major model events on February 11. The Qwen3.5-397B-A17B achieves an 8.6x increase in inference throughput for a 32K context length and set new records for open-source mod

Qwen3.5-397B-A17B, announced February 2026 by Alibaba's Qwen team, is a native vision-language foundation model with 397B total parameters and 17B activated per forward pass, published under Apache License 2.0 on Hugging Face as a post-trained checkpoint compatible with Transformers, vLLM, SGLang, and KTransformers. Th Architecturally, the model builds on Qwen3-Next lineage, combining Gated DeltaNet linear-attention layers with sparse MoE in roughly a 3:1 ratio against full-attention layers, and is trained with multi-token prediction (MTP) across multiple steps. It operates in thinking mode by default, emitting ... blocks before answ