Alibaba Releases Qwen3.5-397B-A17B, Open-Weight Multimodal Model for Agentic AI
On February 16, 2026, Alibaba’s Qwen team released Qwen3.5-397B-A17B, a natively multimodal open-weight MoE model (397B total parameters, 17B active) with strong performance in reasoning, coding, agent capabilities, and vision-language tasks. It features an innovative hybrid architecture for high inference efficiency and a 1M context window in the hosted Plus version.
TLDR
Alibaba released the first model in the Qwen3.5 series on February 16, 2026 (blog dated Feb 15). Qwen3.5-397B-A17B is an open-weight, natively multimodal (vision-language via early fusion) Mixture-of-Experts model. Despite 397 billion total parameters, only 17 billion activate per token thanks to high-sparsity design. It delivers competitive or leading results on agentic, coding, multimodal, and reasoning benchmarks while offering dramatically improved inference throughput (8.6x–19x vs prior generation at common context lengths). The hosted Qwen3.5-Plus variant includes a 1M context window and built-in adaptive tool use.
Key Technical Advances
The model uses a hybrid attention approach combining Gated Delta Networks (linear attention) with sparse MoE for efficiency and capability. Pretraining scaled visual-text tokens, STEM/reasoning data, and multilingual coverage (now 201 languages/dialects, larger 250k vocabulary).
Post-training emphasized broad reinforcement learning across difficult, generalizable environments rather than narrow metric chasing, yielding gains in agentic tool use, planning, and long-horizon tasks.
Benchmarks (per release data) show competitive standing against GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro families across language, vision-language, coding agent, and search agent evaluations, with particular strengths noted in efficiency and certain multimodal/spatial tasks.
Weights are available on Hugging Face, ModelScope, and GitHub under Apache 2.0. The open model is positioned for developers building multimodal autonomous agents.
Why this story matters
Qwen3.5 arrives as part of a concentrated wave of high-capability Chinese open and efficient models released around Lunar New Year. Its combination of frontier-level multimodal agent capabilities, extreme parameter efficiency, open licensing, and claimed cost/performance advantages intensifies global competition on both the capability and accessibility fronts. The timing and technical choices underscore the accelerating shift toward practical, deployable agentic systems rather than raw scale alone.
Sources
- Qwen official blog: “Qwen3.5: Towards Native Multimodal Agents,” released/announced February 16, 2026 (blog timestamp 2026/02/15). https://qwen.ai/blog?id=qwen3.5 (detailed benchmarks, architecture, links to weights and code).
- Reuters coverage: “Alibaba unveils new Qwen3.5 model for ‘agentic AI era’” (Feb 16, 2026), confirming performance and cost claims.
- Hugging Face collection and GitHub repo references in the announcement for Qwen3.5-397B-A17B.
Featured Image Alt Text
Qwen3.5 model release graphic highlighting MoE architecture, multimodal vision-language capabilities, and efficiency gains for agentic workflows
Tags
Qwen3.5, Alibaba, Open Weights, MoE, Multimodal, Agentic AI, Efficiency, China AI