China's chemistry model 3.0 Pro stops answering and starts acting
Two releases out of China this morning point at the same shift: domain models are done proving they can talk about a technical field, and are being rebuilt to do the work inside it.
China's first chemical-industry large model shipped its 3.0 Pro release today, and the pitch has fundamentally changed. The model — built by the Dalian Institute of Chemical Physics (DICP) under the Chinese Academy of Sciences together with iFlytek and Alibaba Cloud — is no longer framed as a question-answering system for chemists. Team lead Ye Mao describes the new version as a move "from knowing things to doing work," and the architecture backs the claim: 3.0 Pro is built as a four-layer stack — model, agent, domain skills and tools, application scenario — in which the agent doesn't just retrieve an answer but decomposes a request, calls simulation software, checks the results and adjusts. When an engineer asks for a process design or an optimization, the model is supposed to work out which physical properties need computing, which tools to invoke, and what to recommend. The scale behind it: more than 400 chemical agents and tools, continued pretraining on a multi-billion-token chemistry corpus, and supervised fine-tuning on billions of tokens of quality Q&A pairs.
The numbers are credible because the benchmark is domain-specific rather than a general leaderboard: 81.96% accuracy on text Q&A and 80.75% on multimodal Q&A, representing relative gains of 20.2% and 31.4% over the previous 3.0 release. What makes this more than a paper exercise is the deployment footprint — over 300 chemical companies, universities and research institutes have registered, with cumulative API calls past 14 million. DICP has also built what it calls China's first petrochemical full-chain data center spanning R&D, design, production and market, and a kiloton-scale intelligent pilot platform is already running in Dalian's Changxing Island to generate the real operating-condition data that neither a lab bench nor a full plant can produce. The stated target is the industry's "valley of death" — the decade-plus slog from lab result to industrial process — compressed into what the team calls a "one step to factory" paradigm. Our read: this is the most serious attempt yet at an agentic industrial model with real equipment attached, and the pilot platform is the part competitors can't copy quickly.
Hisense used the same morning to put an "AI companion" operating system into the living room. JUOS, which the company calls the industry's first household-companion-grade AI OS, runs on Hisense's in-house chips and its Star Sea vertical model, tuned specifically for the multi-person, multi-device, multi-scenario mess that is a real household. Hisense claims over 95% accuracy on family-scenario intent recognition and multimodal first-token latency as low as 175 milliseconds, with a 4+ grade certification from the China Academy of Information and Communications Technology. Three features carry it: an upgraded conversational assistant that handles in-the-moment questions while you watch, a per-person adaptive home screen, and cross-device handoff that links Hisense appliances through its AIoT cloud. Rollout to current Hisense, laser TV and Vidda models begins in September.
The interesting tell here isn't the TV software — it's that a consumer electronics maker is now competing on vertical-model training against general assistants. A household model optimized for "who is asking and what else is happening in this room" is a genuinely different problem from chatbot quality, and Hisense is betting it can own that layer before a frontier lab bothers to.
What to watch: whether any of the joint-lab partners — iFlytek, Alibaba Cloud, Huawei, SUPCON and AVEVA all signed on to a smart-chemicals innovation lab at today's event — ship 3.0 Pro's execution layer into commercial engineering software. That's where "acting" either pays off or doesn't.
Which of these two bets lands first in your work — an agent that runs your simulations, or one that already knows who's in the room? Tell us what you're seeing.
