On February 16, 2026 — Lunar New Year’s Eve — Alibaba released the Qwen3.5 family. The first model out the door, Qwen3.5-Plus, is a 397B-A17B sparse mixture-of-experts, natively multimodal foundation model, with weights open-sourced simultaneously on Hugging Face, GitHub, and ModelScope. Reuters reports the new model is 60% cheaper to use than its predecessor and 8x better at processing large workloads; the official technical notes go further with a concrete figure: 8.6x the decoding throughput of Qwen3-Max on 32,000-token long-context tasks.
The timing tells its own story. Over the past week, ByteDance and Zhipu AI both shipped agent-centered upgrades, and CNBC’s headline cut straight to the judgment: China’s chatbot race is shifting to AI agents. Google DeepMind CEO Demis Hassabis recently assessed Chinese models as only “months” behind their Western rivals. Qwen3.5 is Alibaba’s latest move in that pivot.
Qwen3.5-Plus: A 397B Sparse MoE Catching a Trillion-Parameter Flagship
Qwen3.5-Plus builds on the Qwen3-Next architecture: linear and hybrid attention, a higher-sparsity MoE, multi-token prediction, and stability optimizations. Alibaba claims performance on par with Qwen3-Max — its trillion-plus-parameter flagship — and self-reported benchmarks showing results that rival leading models in language reasoning, code generation, agentic workflows, image and video understanding, and GUI interaction. CNBC duly notes those claims could not be independently verified, but the direction is clear: a sparse design with 397B total and 17B active parameters chasing trillion-scale performance. Lin Junyang, technical lead of Alibaba Cloud’s Qwen team, previewed more open-weight releases over the holiday, and the new RL framework speeds end-to-end training by 3 to 5 times.
Native Multimodal and 201 Languages
Qwen3.5 is natively multimodal: text, image, and video in; text out — trained on trillions of mixed visual and language tokens. The concrete capabilities include understanding videos up to 2 hours long with multi-step temporal reasoning, visual reasoning gains over Qwen3-VL (especially science problem-solving with self-reflection), and visual coding that turns hand-drawn UI sketches into working frontend code. Language coverage expands from 119 languages in Qwen3 to 201 languages and dialects, including low-resource ones like Hawaiian, Fijian, and Niger-Congo family languages.
Agents Are the Battleground: Phones, Desktops, Million-Scale Environments
The center of this release is not chat — it is operation. Qwen3.5’s visual agent autonomously operates smartphones and computers: the mobile agent drives in-app actions from natural language and works across apps, while the desktop agent handles multi-step, long-horizon office tasks. The RL framework natively supports text, multimodal, and multi-turn scenarios, and can absorb agent frameworks and environments at million-scale. CNBC also notes compatibility with open-source agent frameworks like OpenClaw’s — the model layer and the agent layer are converging on each other.
The Open-Source Play Inside China’s Agent Race
Open weights are the established Qwen playbook: Alibaba has open-sourced more than 400 Qwen models, passed 1 billion cumulative downloads, and counts over 200,000 developer-built derivative models. Counterpoint Research analyst Marc Einstein reads it this way: AI firms are preparing for agents to upend traditional Internet business models, “the consequences for those who are not prepared will be severe,” and Chinese firms are well aware — while the 201-language support signals Alibaba’s global ambitions. Put the release on a timeline: a robotics-oriented model one week earlier, an agent flagship on New Year’s Eve — the Chinese holiday has become a launch season for Chinese AI.
What It Means for Developers and Product Teams
Three practical effects. First, a 397B-A17B sparse MoE makes self-hosted inference plausible: with 17B active parameters, the bottleneck shifts toward memory rather than raw compute, and downloadable, fine-tunable open weights give teams that refuse API lock-in a real path. Second, GUI operation and 2-hour video understanding move the agent battlefield from the chat window to the device itself — teams building automation products can evaluate directly whether these capabilities cross the usability threshold. Third, keep your verification habit on benchmark marketing: the gap between self-reported scores and independent evaluation is a recurring pattern, so run your own eval set before committing.
Sources
- Alibaba Open-Sources Qwen3.5, A Natively Multimodal Model — Alibaba Group
- Alibaba unveils new Qwen3.5 model for ‘agentic AI era’ — Reuters
- Alibaba unveils Qwen3.5 as China’s chatbot race shifts to AI agents — CNBC
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.