Alibaba plans a 10T-parameter model and a 500,000-chip cluster
Alibaba's cloud conference in Hangzhou opened this morning with a three-part commitment that is harder to wave off than the usual keynote arithmetic: a model bigger than anything trained so far, the chip to run it, and the power to feed both. Huawei spent the same week putting a coding model inside its own operating system, and a small project on the local-AI forums quietly showed that a model's refusals can be switched off without touching a single weight.
Alibaba says its next Qwen models will reach 5 trillion to 10 trillion parameters — and that its in-house accelerator can wire 500,000 of them into one cluster. At the Apsara Conference in Hangzhou, CEO Eddie Wu framed the company's bet around "machine intelligence": machines are becoming the main producers of thought, and intelligence is turning into a commodity supplied in bulk. His comparison was the industrial revolution turning power into a utility — the steam engine, the combustion engine and electricity only mattered once there were grids to carry them. Wu put the eventual gap at more than 1,000 times the total volume of human thinking, and argued the defining product of the era has not been invented yet: today's coding assistants, in his framing, are the 1882 lightbulb, replacing existing work rather than creating new categories.
The concrete pieces are what matter. The Qwen team reported progress on recursive self-improvement — the loop where a model improves the next version of itself, and the subject of Ezra Klein calls for a ban on self-improving AI. T-Head, Alibaba's chip arm, unveiled the Zhenwu V900, which it says triples the performance of the M890 it showed in May and supports single clusters of up to 500,000 units; the hard part of that number is interconnect, not silicon. Alibaba Cloud, meanwhile, targets more than 20 GW of operated data centres worldwide by 2032, and says customer demand already outruns what it can supply. Read those three together and the strategy is legible: own the model, the chip and the grid, so that no export control and no competitor sits between Alibaba and the tokens it sells. Broadcom can block a licence; it cannot block a power purchase agreement.
Huawei put a coding model inside the HarmonyOS toolchain — and demoed it building a playable game from one sentence. Huawei Cloud launched a HarmonyOS coding large model alongside a CodeArts agent, trained on more than 100,000 pieces of HarmonyOS material and adapted to ArkTS and HarmonyOS development conventions. Huawei's numbers, not ours: errors per thousand lines of code down by more than 80%, first-compile pass rate up by 78%, token spend per task down by 20%. The demo is the part worth watching — an ArkTS side-scrolling platformer with gesture controls, built and installed in about 20 minutes, one control bug found and self-fixed in roughly three, then a checkpoint feature added in four more minutes without breaking what already shipped. The agent covers requirements through compilation and can drive the on-device HarmonyOS emulator, which Huawei calls an industry first and which is really the point: whoever owns the emulator owns the feedback loop.
A community project can now turn an LLM's refusals off without editing a single weight. A developer released phantom-kv, a roughly 18 MB bank of learned key/value tensors that loads into the model's KV cache as context; off means unloaded, and nothing on disk changes. Notably, the author reports that a cruder split — using only the first half of the tensor bank — improved general behaviour but did nothing to refusals, because the refusal signal lives in the deeper layers. That single negative result is the most useful thing in the release: it maps where in a model's depth a safety behaviour actually sits, and it hands anyone with download rights a switch that ships separately from the model.
What to watch: whether Alibaba's parameter target arrives with a training bill it can actually finance, and whether Huawei's coding model shows up in developer numbers rather than keynote demos.
Would you ship an agent that writes and compiles its own app features before you've reviewed the last one? Tell us in the comments.
