Meituan's LongCat-2.5 went up on Hugging Face six days ago under MIT. As of this morning there are 1,912 derivative repositories, and the interesting finding is not the count. It is that the best-performing derivatives beat the base model on the very benchmark the base model was announced with.

What shipped

LongCat-2.5 is a native multimodal mixture-of-experts with 1.6 trillion total parameters and about 48 billion active, trained end to end on domestic accelerators. It scored 88.4 on Terminal Bench v2.1 at release, within half a point of GPT-5.6-Sol and Claude Opus 5. MIT licensing, weights plus inference code.

The active-parameter count is what made the derivatives possible. A 48B-active model fits on eight 80GB accelerators with room for a LoRA, which means a lab with a modest cluster can fine-tune it. A dense 1.6T model would have been read-only for everyone outside a hyperscaler.

Three kinds of derivative, only one of which is useful

The first kind is a merge: two or three community LoRAs combined with a weighted average, no new training. These are cheap, numerous, and mostly noise. The second is a domain tune — legal, medical coding, a specific language pair — trained on a few thousand curated examples. That is where the value is. The third is a quantized-plus-system-prompt repackaging that presents itself as a new model; it is not, and the model card usually admits it if you read to the bottom.

Of the 1,912 repositories, roughly 240 include an evaluation table with numbers on a named public benchmark. That is the filter worth applying before anything else.

The one that matters

A two-person team in Bengaluru published a coding tune eleven hours after the base weights landed, trained on 18,000 filtered pull-request diffs, and it scores 90.5 on Terminal Bench v2.1 — 2.1 points above base. It also regresses 4 points on long-context retrieval, which the model card states plainly. That trade is normal, and the disclosure is what makes it usable.

The licence nuance nobody reads

MIT on weights is not MIT on everything. The training data is undisclosed, and at least three derivatives advertise themselves as "commercially safe" with no basis beyond the weight licence. If your legal team needs a data-provenance answer, open weights cannot give it to you. That is a property of the whole open-weight ecosystem rather than a LongCat problem, but the MIT label makes it easier to forget.

How to pick one

Four checks, in order. Does the model card say what data the tune used and what it regressed on? Is there an eval table with a named harness, or only a screenshot of a chat window? Can you reproduce that eval in an afternoon on your own tasks? And does the repository have a commit in the last two weeks — an abandoned tune on a live base model is a support burden with no upside.

The base model is good. The derivatives are the reason to care, and the ones worth using will tell you exactly what they broke.