Qwen3.5 397B-A17B

Alibaba's Qwen team officially released Qwen3.5-397B-A17B, the first open-weight model in the Qwen3.5 series. It is a native vision-language model delivering strong results across reasoning, coding, agent, and multimodal benchmarks. The hybrid architecture fuses linear attention via Gated Delta Networks with a sparse mixture-of-experts, running 397 billion total parameters with 17 billion activated per forward pass for inference efficiency. Language and dialect support expanded from 119 to 201, broadening global accessibility. On Qwen3.5-397B-A17B's leaderboard evaluations, the model posts 87.8 on MMLU-Pro, 88.4 on GPQA, 76.4 on SWE-bench Verified, 79.0 on MMMU-Pro, 72.9 on the Berkeley Function Calling Leaderboard, 95.6 on τ²-bench Telecom, and 93.33 on AIME 2026, with an AA-Omniscience index of -30 signaling notable hallucination weakness. The companion Qwen3.5-Plus is a separate hosted variant on Alibaba Cloud Model Studio with a 1M-token context and built-in adaptive tools.

evals.report independently tracks 22 reported benchmark scores for Qwen3.5-397B-A17B released on February 16, 2026, with per-benchmark source-status labels (Official, Verified, or Unverified). Coverage spans SWE-bench Verified, GPQA Diamond, Humanity's Last Exam, Berkeley Function Calling Leaderboard, MMMU-Pro, Artificial Analysis and Epoch Capabilities indices, τ²-bench Telecom, AIME 2026, GDPval, and WebDev Arena. No single combined leaderboard score is published, reflecting a per-benchmark rather than aggregated view. Highlights from the tracker include SWE-bench Verified at 76.4 percent, GPQA Diamond at 88.4 percent, MMMU-Pro at 79.0 percent, AIME 2026 at 93.33 percent, WebDev Arena at 1393 Elo, and an Epoch Capabilities Index of 146.1. The AA-Omniscience score of -30 stands out as a notable hallucination weakness for the variant. The page also flags many scores as Unverified, signaling that independent reproduction is still pending for several measures.

Alibaba announced the open-source release of Qwen3.5 on February 16, 2026, with the first model in the series being Qwen3.5-397B-A17B, also branded as the hosted "Qwen3.5-Plus". The model is described as a natively multimodal foundation model trained from scratch on trillions of vision-language tokens spanning multilingual text, images, videos, STEM, and reasoning data, with support for 201 languages and dialects, up from 119 in the Qwen3 series, including low-resource languages such as Hawaiian, Fijian, and Niger-Congo languages. According to Alibaba's announcement, Qwen3.5-397B-A17B targets strong performance across language understanding and reasoning, code generation, agentic workflows, image and video comprehension, and GUI interaction, with results rivaling leading frontier models in versatility and capability. A central design priority is inference efficiency: the model pairs a linear attention mechanism with a sparse mixture-of-experts (MoE) architecture to lower compute requirements while remaining suitable for broad real-world deployment of multimodal agentic applications.

On February 16, 2026, Alibaba officially open-sourced Qwen3.5, with the headline release being the Qwen3.5-397B-A17B variant (also branded "Qwen3.5-Plus"). The model is described as a natively multimodal foundation model that combines a linear attention mechanism with a sparse Mixture-of-Experts (MoE) design, activatin The Qwen3.5-397B-A17B was trained on trillions of vision-language tokens spanning multilingual text, images, videos, STEM, and reasoning data, supports 201 languages and dialects (up from 119 in Qwen3), and processes text, image, and video while generating text. According to Alibaba's announcement, it delivers strong p

The Baidu encyclopedia entry for Qwen3.5 names both Qwen3.5-Plus and Qwen3.5-397B-A17B as versions in Alibaba's February 16, 2026 flagship release and describes the underlying architecture as a Hybrid Attention Mechanism combined with a Sparse Mixture-of-Experts design optimized for logical reasoning, mathematical comp The same Baidu entry documents rapid ecosystem adoption following the February 16 launch, noting that by February 18, 2026, international hardware and framework vendors including NVIDIA, AMD, and Apple had completed adaptation for Qwen3.5, while domestic Chinese GPU and platform providers such as Huawei Ascend, Moore T

Qwen3.5-397B-A17B is the first open-weights release in Alibaba's Qwen3.5 series, announced by the Qwen team as a native vision-language foundation model with 397B total parameters and 17B activated per forward pass, per the Alibaba Cloud "Qwen3.5: Towards Native Multimodal Agents" post. The Qwen blog dates the official The model is positioned for reasoning, coding, agent tasks and multimodal understanding, with early-fusion training achieving parity with Qwen3 while outperforming Qwen3-VL on reasoning, coding, agents and visual understanding, per the Hugging Face model card. Independent evaluation placed it at 45 on the Artificial An