Meituan's LongCat team published the LongCat-2.5 weights on November 16 under the MIT licence. Two facts make it more than another open model release: the parameter count, and what it was trained on.

The model

LongCat-2.5 is a native multimodal mixture-of-experts with 1.6 trillion total parameters and roughly 48 billion activated per token. It is aimed at long-horizon agent work rather than chat, which shows up in how the team reports results: terminal tasks, not conversation benchmarks.

Trained without Nvidia

Every training run was done on domestic AI accelerators. The team credits three techniques for making that viable: supernode-level parallelism, sharding, and overlapping computation with communication. Long-sequence supervised fine-tuning reached 86% higher per-device throughput than a distributed data-parallel baseline.

The self-hosting reference configuration uses 194 Ascend chips. That number matters more than the parameter count for anyone outside China who is wondering whether the stack is reproducible: it is a specific, quotable cluster size, not a vague claim about "domestic compute".

The benchmark that counts

On Terminal Bench v2.1, LongCat-2.5 scores 88.4. GPT-5.6-Sol scores 88.8 and Claude Opus 5 scores 89.1. Being within half a point of both is the strongest claim in the release, because terminal benchmarks reward exactly what agent work needs: reading errors, running commands, recovering from a failed step.

Treat it as one data point. Terminal Bench is narrow, and every lab tunes for the benchmarks it quotes. But 88.4 on an open MIT-licensed model is not a rounding error.

Why MIT matters here

Apache and MIT licences are what let a model enter a commercial product without a legal review that stalls for a quarter. A 1.6-trillion-parameter model under MIT is still impractical to serve for most teams — 194 accelerators is not a hobby cluster — but it is entirely practical for a cloud provider or a large enterprise to host internally, and that is where open weights actually get used.

It also means the model can be fine-tuned and the result kept private. For companies in regulated industries, that is usually the deciding factor.

The honest caveats

A 48-billion-active model at this size will be slow and expensive per token compared with a 12-billion-active model, so the economics only work when the task genuinely needs the capability. Long-horizon agent runs qualify; a customer-support bot does not.

There is also no published evaluation of long-context behaviour beyond the headline window, and no independent reproduction of the training claims yet. Both will arrive within weeks if the release is real.

How to evaluate it

If you have the hardware or a provider that hosts it, the useful test is a multi-step terminal task that fails halfway: install something that breaks, then ask the agent to fix the environment. Compare how many turns it needs against your current model. Long-horizon agents live or die on recovery, not on the first answer, and that is precisely what the benchmark is measuring.