LongCat-2.0 - 1.6T MoE trained entirely on AI ASICs

Hi everyone!

LongCat-2.0 is a 1.6T-parameter MoE model with about 48B active parameters per token, 1M context, and open weights under MIT.

But it was not trained in the usual Nvidia-heavy way. The full training run was built on AI ASIC superpods, over more than 35T tokens, with no rollback or irrecoverable loss spike.

Training a trillion-scale model is already hard. Getting that run stable on alternative hardware is probably the more interesting story here 🤔

Curious to see how you frame the model strengths beyond the short tagline.

the "no rollback, no irrecoverable loss spike" claim on ASIC superpods is the part that would make me trust this more than the raw parameter count. that's usually where alternative-hardware training runs quietly fall apart, not in the benchmarks people see later. one thing I'd want to know before betting an agent pipeline on this: since it integrates with Claude Code and OpenClaw, did the ASIC training path change the numeric format at all versus a typical bf16 GPU run, and if so does that show up as any weirdness once you quantize down for inference, or does it behave like any other MoE checkpoint at that point

awww, cuteeee

Congrats to the LongCat team!

As someone based in China, Meituan is a household name to me — I use it almost every week for food delivery and local services. So it’s honestly a bit surprising, in a good way, to see the same company open-source a 1.6T MoE model under MIT. That contrast makes this launch especially interesting.

I’m also exploring AI-assisted PM workflows and building my own agent skills, so the combination of open weights, 1M context, and post-training for coding/agentic workflows really stands out. Curious if the team has any best practices for using LongCat-2.0 in long-horizon agent workflows where the context includes product docs, code, and task history.

How does the 560B MoE actually handle latency in real-world agent loops, and is there a hosted endpoint or do we need to self-host the whole thing to use it?

Curious how the 560B MoE actually handles rate limits in practice, like does Meituan expose a hosted API or are we expected to self-host the weights for any serious throughput?

Tried LongCat-Flash-Thinking on a multi-step coding prompt and was surprised how fast it reasoned through edge cases without losing track. Solid open-source move from Meituan.

How does LongCat-Flash-Thinking handle context length on really long documents, and is it open weights or just open source code?

How does the 560B MoE setup actually translate to faster inference in practice compared to dense models of similar size, and is it open weights under a permissive license?