LongCat-2.0

How does LongCat decide which parts of a conversation are most important to retain over time?

The stable run on ASIC superpods with no irrecoverable spike is the genuinely impressive part here, more than the parameter count. For agent use the metric I care about is long-horizon adherence rather than single-turn tool accuracy: in our loops the model at tool call 30 has usually forgotten a constraint it agreed to at call 5, and 1M context helps recall without stopping that instruction decay. Did the agentic post-training target staying on a plan across many tool calls, or mostly one-shot tool-call correctness?

How does the 560B MoE setup actually perform on smaller hardware for fine tuning, or is it really only practical to run through Meituan's own infrastructure?

how does the 560B MoE setup hold up on longer context tasks compared to dense models like DeepSeek?

Curious how the 560B MoE setup actually feels in latency for real time stuff like chat compared to something like DeepSeek, and what kind of hardware you'd realistically need to run a distilled version locally?

Runs surprisingly snappy for a 560B MoE — I had low expectations but the reasoning responses came back fast and actually thought through edge cases instead of hand-waving.

the reasoning speed on long context prompts genuinely surprised me, felt closer to a smaller model than a 560B MoE.

Curious how this stacks up against other open-source MoE models on benchmarks like HumanEval or MATH, and is it actually free to deploy commercially or are there usage limits built in?

how does the 560B MoE setup handle latency on longer reasoning chains, and is there a hosted endpoint or is it self-host only?

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 🤔