Qwen's agent built a research-grade telescope simulator in 3 days
Two stories about the layer underneath the models: the agent that now plans the observation, and the memory the whole stack is short of.
Alibaba says its Qwen Office agent helped a research team at the National Astronomical Observatory of China build a large-aperture, research-grade telescope simulation system in three days, for under 1,000 yuan — work that previously went to an outside software vendor, took about three months and cost tens of thousands of yuan. The system packages the telescope's components, sensor states and observing conditions as standard MCP interfaces, so an agent can monitor and call them. Two scientific models sit behind those interfaces: a cross-survey alignment model that scans public survey data for very early supernova candidates — telling the agent what is worth looking at — and a short-term local weather model that weighs cloud cover, wind speed and humidity to judge whether the night is usable. From there the agent closes the loop on its own: sense state, rank targets, generate an observing plan, validate the flow before it reaches real hardware.
It is already wired into the observatory's "Sitian" pathfinder and prototype telescopes, according to Alibaba. The company says the agent has flagged eight very-early supernova candidates, two of which triggered follow-up observations when weather allowed, and that the simulation will generate training data for a telescope-control vision-language-action model. The honest framing is that those eight are candidates surfaced by the system, not confirmed discoveries, and there is no paper attached to the deployment — it is a company announcement. The declarative number is the three days; the durable asset is the interface layer. Once an instrument's parts and sensors are exposed as callable tools, the same pattern ports to any telescope, and the vendor quote for a sim environment stops being the gate. We have seen this shape before from Qwen agents doing work humans used to supervise — last week an agent retrained its own model after being asked to fix a bug.
Reuters reports that CXMT, China's largest DRAM maker, is preparing to expand into NAND flash memory — setting up an R&D production line at its new Beijing plant and a research institute in the capital with NAND development among its projects. That puts it against Samsung and SK hynix in the market they have dominated, and against domestic rival YMTC, which has held the flash side for China. Tight memory supply is the reason timing is on the table: the AI buildout has given smaller Chinese suppliers real pricing leverage at home, even though CXMT and YMTC still trail the majors and remain more exposed to lower-priced products. CXMT raised RMB 57.92 billion — about $8.6 billion — in Asia's largest IPO this year, is planning a second Beijing memory plant, and has been in funding talks with a local government-backed manufacturing hub, according to the same reporting.
What the reporting does not settle is scale. Reuters sources could not say when the R&D line starts operating, or whether CXMT intends to move past research and trial runs into commercial NAND production — and an R&D line is not a market entry. The memory squeeze that has already moved Chinese AI chip pricing — China's AI chip prices jump 50% as the memory shortage bites — is the demand signal CXMT is reading. Whether it converts into supply depends on a lane that has not started running yet.
What to watch: whether Sitian's observations produce a published result from the candidate list, and whether CXMT's Beijing NAND line moves beyond R&D within the next two quarters.
Instruments that plan their own observing night are a small step with an obvious end state — so is the interesting part the agent, or the standardised interface that let it in? Tell us in the comments.
Sources: 雷峰网 Leiphone · Sina Finance (IT之家) · DoNews · Reuters — China's CXMT eyes flash-memory push · TrendForce · DigiTimes
