LongCat-2.5 Open Weights: MIT License, NPU Docs Only

Meituan put the weights for LongCat-2.5 on Hugging Face on November 9, about a week after it launched the model as a hosted service, according to VentureBeat. The license is MIT. Anyone can download the checkpoint, change it and use it commercially.

Then you scroll to the Deployment section of the model card, and it’s one line long. “LongCat-2.5 can be deployed on NPU platforms,” it says, with a link to Meituan’s SGLang-FluentLLM project. No GPU instructions appear on the card.

My read: “open weights” and “easy to run” are different claims.

This release is a clean example of the gap.

What the model is, per Meituan

The model card describes a native multimodal mixture-of-experts model for long-horizon agent work, with 1.6 trillion total parameters and about 48 billion activated per token. It builds on LongCat-2.0 and adds image and video understanding to agentic coding.

The license grants no rights to Meituan’s trademarks or patents, according to the same card. My read: that’s a common MIT-style carve-out, but it matters if you plan to build a product around the name.

Meituan’s own tech blog says the model was trained and tuned for inference on “domestic AI accelerators.” It reports that a 4-bit scheme (W4A4) on those chips raised prefill throughput by 43.3% and decode throughput by 42.77% against an 8-bit baseline. Those are Meituan’s online stress tests, not anyone else’s.

The size question

The card says 1.6 trillion parameters. Hugging Face’s own metadata for the repository counts 1,779,752,648,320, or about 1.78 trillion, almost all of it stored in BF16. I read that figure from the Hugging Face API, so it comes from the files and not from a press write-up.

At two bytes per BF16 parameter, my arithmetic puts the download near 3.6 terabytes before you’ve loaded anything. The 4-bit version Meituan describes would be much smaller, but the card links no separate quantized checkpoint. I couldn’t confirm whether one exists.

So the practical reading, and this is my inference, is that most US teams would reach LongCat-2.5 through Meituan’s hosted API or a third-party host. Running it yourself means a multi-node cluster and software support the card doesn’t spell out for GPUs.

Benchmarks, Meituan’s version

Meituan compares the model against GPT-5.6-Sol and Claude Opus 5, among others. Scores with an asterisk on its table are taken from other labs’ reports. Unmarked ones were run by Meituan and are labeled “indicative only.” Here are four rows.

Benchmark	LongCat-2.5	Claude Opus 5	GPT-5.6-Sol
Terminal Bench v2.1	88.4	89.1	88.8
Terminal Bench v4.0	38.5	53.9	37.3
DeepSWE v1.1	64.6	68.8	72.7
BrowseComp	91.9	90.8	90.4

It wins one row, loses most of the rest, and the jump from LongCat-2.0 on Terminal Bench v4.0 (3.0 to 38.5) is the part that deserves a second look. Meituan printed the losses next to the wins, which I’d rather see than not. Still, every number here comes from the company that built the model, with test settings it chose.

The policy shadow

VentureBeat noted that the Department of Homeland Security ordered civilian agencies in October to drop DeepSeek, and that it found no US federal restriction naming Meituan or LongCat-2.5 as of November 9. We covered that order in our post on the DHS DeepSeek ban. The DeepSeek order names DeepSeek and its affiliates only.

That is a snapshot, not a promise.

A company that sells to federal agencies should check its contracts before building on any Chinese-origin model, whatever the license says.

One more oddity: the Meituan blog post still lists its GitHub link as “coming soon,” while the weights are already out. I’d expect more deployment detail to arrive, though nothing I read says when.

Sources

Note on sources: the parameter counts and deployment wording come from the model card and repository metadata. The license-date and DHS context come from VentureBeat. Every benchmark figure is from Meituan’s own table.