PRODUCT • • 4 min read

Meituan Open-Sources LongCat-2.5, a 1.6T MoE With About 48B Active

Thumbnail for: Meituan Open-Sources LongCat-2.5, a 1.6T MoE With About 48B Active

Meituan said on November 10 that LongCat-2.5 is open source under the MIT license, including commercial inference and further training, with no extra grant required. The company and the Hugging Face model card both describe a native multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token. The Hugging Face repo was created on November 4. Weight shards were uploading on November 4 and November 5, and the newest shard commit is 02:48 UTC on November 5. Model-card edits continue through November 10. The Decoder reported the release on November 15.

The numbers

  • 1.6 trillion total parameters, about 48 billion active per token. Meituan's November 10 post and the Hugging Face card both use those figures.
  • MIT license. The LICENSE file on the repo is the standard MIT text, copyright 2026 Meituan. The November 10 post says commercial inference and further training need no extra grant.
  • Under 80GB per device. Meituan says training ran on domestic AI accelerators at that memory limit, with FSDP throughput 86% above a DDP baseline.
  • About 1.78 trillion parameters in the Hugging Face safetensors metadata. Meituan's description says 1.6 trillion. The posts do not explain the gap.

What Meituan measured

The model card lists a Terminal Bench v2.1 score of 88.4 for LongCat-2.5, with no asterisk. Meituan says unmarked scores are in-house and indicative only. It lists Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8, and DeepSeek-V4.1-Flash at 90.6, and marks those three with an asterisk. Asterisks, the card says, are cited from public reports, blogs, or leaderboards, not from the same in-house run. DeepSWE v1.1 moves from 9.7 for LongCat-2.0 to 64.6 for LongCat-2.5. Both of those cells are unmarked. The Decoder repeats the 88.4 score and the DeepSWE jump.

Meituan's November 10 post says the model supports a one million token context. The English card's evaluation notes use a one million token cap on DeepSWE and a 512,000 token cap on Terminal Bench. Those lines are test settings. The product claim of a one million token context is in the Chinese post. The November 1 LongCat blog introduces the same 1.6 trillion and 48 billion figures and the work on domestic AI accelerators. It is an introduction, not the weights announcement.

On inference, the card says expert weights and activations use W4A4 quantization so domestic accelerators can run native INT4 math. Against a W8A8 mixture-of-experts baseline, Meituan reports 43.3% higher prefill throughput and 42.77% higher decode throughput, and says accuracy stayed comparable on the benchmarks it evaluated. The card says deployment is on NPU platforms and points to SGLang-FluentLLM. It does not name Nvidia. The Decoder's headline says the model never touched an Nvidia GPU, and it describes a sample deployment on 194 Huawei Ascend chips. That chip count and the Nvidia exclusion are The Decoder's, not a sentence in the Meituan post or the model card.

Hugging Face lists the repo under the image-text-to-text pipeline, and the card says thinking mode can be on or off. Meituan's November 10 post says visual signals are stored as discrete tokens and handled by the same backbone as text. That is the mechanism it cites for visual coding and 3D work. The post describes the capability. It does not attach a customer count or a price.

What we are watching

The MIT license, the 1.6 trillion and 48 billion figures, and the November 4 to November 5 shard uploads match Meituan and the Hugging Face history. The 1.78 trillion file count does not match the 1.6 trillion sentence, and these posts do not say why. The benchmark table mixes in-house scores with cited scores. A reader who wants the chip vendor, the size of the training cluster, or a third-party rerun of Terminal Bench will not find them in the model card or the November 10 post. The Decoder supplies a Huawei figure the primary posts do not.

Quick answers

What happened with Meituan?
Meituan said on November 10 that LongCat-2.5 is open source under the MIT license, including commercial inference and further training, with no extra grant required. The company and the Hugging Face model card both describe a native multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token. The Hugging Face repo was created on November 4.
What are the key numbers?
1.6 trillion total parameters, about 48 billion active per token. Meituan's November 10 post and the Hugging Face card both use those figures.
What happens next?
The MIT license, the 1.6 trillion and 48 billion figures, and the November 4 to November 5 shard uploads match Meituan and the Hugging Face history. The 1.78 trillion file count does not match the 1.6 trillion sentence, and these posts do not say why. The benchmark table mixes in-house scores with cited scores.

This article was ultrathought.

Pass it on

The story stays open.

The story stays open. Share it once, or start the free brief, to open the sample format.

Sample of the morning brief

Not a live send. The real brief arrives only after you confirm your email.

The example stays hidden until you share or start the brief.

Founding reader means 3 people confirm the brief from your link. The count lives on the subscriber record. This page does not show a live tally.

Related stories

More on Hugging FaceMore AI Models
Meituan Open-Sources LongCat-2.5, a 1.6T MoE With About 48B Active | Ultrathink