Summary: Alibaba released Qwen3.5-Plus on February 16, 2026 as an open-weight, natively multimodal 397B-A17B sparse mixture-of-experts model aimed at agent workflows across phones, desktops, and long videos. Reuters reported 60% lower running cost and 8x better large-workload performance, while CNBC noted that Alibaba’s benchmark claims were not independently verified.
Lunar New Year’s Eve — February 16, 2026 — was the date Alibaba picked to roll out the Qwen3.5 family. First out is Qwen3.5-Plus, a natively multimodal foundation model built on a 397B-A17B sparse mixture-of-experts design, with open weights landing on Hugging Face, GitHub, and ModelScope the same day. Reuters frames the numbers this way: running the new model costs 60% less than its predecessor, and it copes with large workloads 8x better; the official technical notes add a sharper figure — 8.6x the decoding throughput of Qwen3-Max on 32,000-token long-context tasks.
The calendar date is part of the message. Inside a single week, ByteDance and Zhipu AI each pushed agent-centered upgrades out the door, and CNBC’s headline delivered the verdict without hedging: China’s chatbot race is shifting to AI agents. Google DeepMind CEO Demis Hassabis recently put the gap between Chinese models and their Western rivals at just “months”. With Qwen3.5, Alibaba plays its newest card in that pivot.
Qwen3.5-Plus: Chasing a Trillion-Parameter Flagship with a 397B Sparse MoE
The bones come from the Qwen3-Next architecture: linear and hybrid attention, a sparser MoE, multi-token prediction, and stability optimizations. Alibaba claims the result performs on par with Qwen3-Max, its flagship past the 1-trillion-parameter mark, and submits self-reported benchmarks that rival leading models across language reasoning, code generation, agentic workflows, image and video understanding, and GUI interaction. CNBC duly flags that none of it has been independently verified — still, the direction is unmistakable: 397B total parameters and 17B active ones, arranged sparsely, in pursuit of trillion-scale results. Lin Junyang, who leads the technical side of Alibaba Cloud’s Qwen team, teased further open-weight releases over the holiday, and end-to-end training runs 3 to 5 times faster on the new RL framework.
Natively Multimodal, Covering 201 Languages
Multimodality here is native, not bolted on: text, image, and video go in; text comes out; and training ran on trillions of mixed visual and language tokens. The capability list includes following videos that run up to 2 hours while doing multi-step temporal reasoning, visual reasoning that improves on Qwen3-VL — most notably science problem-solving paired with self-reflection — and visual coding, where a hand-drawn UI sketch becomes working frontend code. Language coverage stretches from 119 languages in the Qwen3 era to 201 languages and dialects, folding in low-resource entries such as Hawaiian, Fijian, and languages of the Niger-Congo family.
The Real Battleground Is Agents: Phones, Desktops, and Million-Scale Environments
Chat is not what this release is about; operation is. Qwen3.5’s visual agent takes over smartphones and computers on its own: on mobile, natural-language commands drive in-app actions and coordination across apps; on the desktop, it carries office work that spans many steps and long horizons. The RL framework natively covers text, multimodal, and multi-turn scenarios and is built to take on agent frameworks and environments at million-scale. CNBC adds that it interoperates with open-source agent frameworks like OpenClaw’s — more evidence that the model layer and the agent layer are converging.
Open Weights Inside China’s Agent Race
Giving the weights away is long-standing Qwen procedure: more than 400 Qwen models open-sourced so far, cumulative downloads past 1 billion, and over 200,000 derivative models built by developers. Counterpoint Research analyst Marc Einstein reads the field this way: AI firms are bracing for agents to upend traditional Internet business models — “the consequences for those who are not prepared will be severe” — and Chinese firms clearly know it; the 201-language support, meanwhile, telegraphs Alibaba’s global ambitions. Drop the release onto a timeline — a robotics-oriented model arrived one week earlier, the agent flagship landed on New Year’s Eve — and the Chinese holiday has turned into launch season for Chinese AI.
Why Developers and Product Teams Should Care
Three effects worth flagging. First, a 397B-A17B sparse MoE puts self-hosted inference within reach: with only 17B active parameters, the cost bottleneck sits in memory rather than raw compute, and downloadable, fine-tunable, self-deployable open weights hand teams that want no part of API lock-in a genuine alternative. Second, GUI operation and 2-hour video understanding push the agent battlefield off the chat window and onto the device itself, so teams building automation products can judge directly whether these capabilities clear the usability bar. Third, keep verifying benchmark marketing: self-reported scores and independent evaluations diverge often enough that running your own eval set before committing is the safer default.