iFlytek ships Spark X2.5, trained end-to-end on Chinese silicon
Two Chinese releases this morning bracket the same bet — that models and agent products can be built without American chips — plus a security firm's answer to the coding-agent mess we covered overnight.
iFlytek released Spark X2.5, a 293B-parameter mixture-of-experts model with about 30B active, and the headline claim is not the size — it's the hardware. The company says the model was trained and is served entirely on domestically produced compute, continuing work on what it calls the hard problems of ultra-long-sequence training and inference on Chinese accelerators. The release focuses on coding and agent capability while also improving math and general comprehension, and it is live now on iFlytek's open platform. The smaller edge models in the same family — 4B and 1.7B — were open-sourced ahead of this launch and handle up to a million tokens of context, aimed at local deployment for long documents, code assistance, and data analysis. With the flagship out, iFlytek now has a matched cloud-and-edge line.
The significance is narrower than "China catches up" and more useful than that. A 293B MoE that never touches an Nvidia cluster is a proof point for the whole domestic-stack thesis, and it lands the same week reporting showed Chinese open-weight models out-drawing American ones on OpenRouter for the nineteenth straight week. It also arrives days after fresh reporting on how easily export controls are routed around — so the counterfactual matters: if a lab can train a model this size on Chinese silicon, the leverage of chip restrictions shifts from "can they build at all" to "how fast and how many." Treat the domestic-compute claim as iFlytek's own; no independent benchmark of X2.5 has landed yet, and the company has not published full eval numbers in the coverage so far.
Alibaba's Qwen Office (千问办公) shipped what it calls the industry's first multi-person workspace, and it quietly changes what an "AI office tool" is competing with. Describe a workflow in plain language and the product generates a publishable web app that up to a hundred people can use at once — with role-based permissions, a cloud database, an admin backend, and one-click publishing, which is the part existing AI workbench tools lack (they mostly produce single-user pages: a résumé, a portfolio, a sign-in sheet). Alibaba's own examples are the tell: a market organizer processing a hundred vendor applications and deposits, a school workspace where each parent sees only their own child's grades, and an enterprise board tracking dozens of influencer campaigns or store openings. The pitch against SaaS is explicit — a commerce suite can run tens of thousands of yuan a year in subscription fees, and custom builds hundreds of thousands, with months of lead time.
Scale is the reason to take it seriously: Qwen Office passed 30 million users in its first month, and more than half of those are enterprise users, per Chinese financial outlets — numbers Alibaba released and that no one has independently audited. The strategy underneath is the interesting part. Most agent products are sold as personal productivity, where the ceiling is an individual's afternoon. Multi-person workspaces put the agent inside an organization's actual process, holding the data, the permissions, and the state — which is a much stickier place to be, and a much harder one to get right.
Trail of Bits published coop, a Rust tool that runs Claude Code and Codex inside disposable virtual machines instead of on your laptop. The agents get everything they need — Docker, git, compilers, package managers — inside a VM that is isolated, reproducible, and cheap to create and destroy, so a run that goes wrong costs you a machine you were going to throw away. It's the obvious response to the class of problem in this morning's other story: we covered how merely opening a repository can execute code through your coding agent in A repo you only opened can run code through your AI coding agent — eight findings across seven agents, four still unpatched. Application-level sandboxes keep getting bypassed because they inherit assumptions from the host; a VM boundary is blunter and harder to talk your way out of.
What to watch: whether X2.5 gets an independent benchmark run, and whether multi-user agent workspaces from Chinese vendors get any real enterprise adoption outside China.
Would you run your coding agent inside a disposable VM if it meant giving up filesystem access to the rest of your machine? Tell us in the comments.
Sources: coop (GitHub)
