|
A thing that is making headlines today is the announcement from Meituan that it had trained a 1.6-trillion-parameter LLM entirely on China-made chips. “LongCat-2.0” is being pitched as a particularly strong agentic coding model, with impressive benchmark results, and is being discussed as a milestone for both the Chinese side of the AI race and for the open source/open weights community.
On its face, the news is impressive, and might suggest almost a DeepSeek-like moment. The 1.6-trillion-parameter aspect is striking – Meta’s own unreleased Behemoth model (discussed in yesterday’s newsletter) is thought to be around 2-trillion parameters, and that’s at the cutting edge. But there are a few reasons to take the announcement with a grain of salt:
It’s a Mixture-of-Experts model. Most models, like the ones powering Claude 4.8 or GPT-5.5, cram all their parameters into a single big “brain”; but a MoE model essentially comprises a bunch of smaller brains that are experts in different areas. So LongCat-2.0 isn’t activating all 1.6T parameters for a given prompt, and is instead activating one of its 48-billion-parameter brains per token.
Meituan hasn’t actually said which chips its model was trained on. It said it was trained on a domestic, 50,000-chip cluster, but not actually saying whose chips were involved is odd. Surely the chip partner would be proud to be involved in this project?
While Meituan is framing LongCat-2.0 as an open source project, it’s… not. Not yet, at least. Its page on Hugging Face – a widely used hub for open source code – currently says “Model weights coming soon”, as does its GitHub repository. That could indeed arrive sooner or later, but it’s an odd omission at the time of launch.
Who the heck is Meituan? Outside of China, it’s a little-known company, at least in comparison to the more established AI players like Alibaba, ByteDance, and DeepSeek. To be fair, in China, it’s a huge brand – but it’s known for things like booking hotels and food deliveries. It only launched its LongCat AI arm in 2023, so it’s kind of surprising that it has made such progress since then.
None of this is particularly damning, and I don’t want to be dismissive, but I have also come to be a little bit skeptical of the claims of China-based open source/open weights competitors. There have been a few big model launches in recent years with staggering claims about performance on widely used benchmark tests, but when used in practice, for real work, they have disappointed not just me but many others – suggesting, in those cases, that they might have been trained specifically to beat the benchmark tests, rather than built for general intelligence.
As always, time will tell. But I want to give the last word here to Peter Wildeford, Head of Policy at the AI Policy Network, who had this to say on X, the everything app:
|
