Timeline

Meituan releases LongCat-2.5 open weights

A 1.6-trillion-parameter multimodal mixture-of-experts model with about 48 billion active parameters, under an MIT licence, with 4-bit quantisation aimed at domestic Chinese accelerators.

  • Models & capabilities
  • Open weights & ecosystem
  • Minor

Meituan’s LongCat team published the weights of LongCat-2.5 on Hugging Face under an MIT licence, which the model card says grants no rights to Meituan trademarks or patents. Hugging Face’s repository history shows the weight files uploaded late on 4 November UTC, which is 5 November in Beijing; the card itself carries no release date.

The card describes a “native multimodal MoE model for long-horizon agentic tasks”: 1.6 trillion total parameters, about 48 billion active per token, with image and text input and a switchable thinking mode. It builds on the team’s earlier LongCat-Next work.

Meituan’s benchmark table compares it with models including GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5, Kimi K3 and DeepSeek-V4.1-Flash. The card says unmarked scores were measured in-house and are “indicative only”, while starred scores come from public reports or leaderboards. The results are mixed. LongCat-2.5 leads the listed models on BrowseComp, at 91.9 against 91.5 for DeepSeek-V4.1-Flash, but trails on others: 88.4 on Terminal Bench v2.1 against 89.1 for Claude Opus 5 and 90.6 for DeepSeek-V4.1-Flash, and 38.5 on Terminal Bench v4.0 against 53.9 for Claude Opus 5. No independent evaluation was cited.

A large part of the card concerns deployment. Meituan says it applied W4A4 quantisation, 4-bit weights and activations, to the expert layers for “native INT4 computation on domestic AI accelerators”, and reports roughly 43% higher prefill and decode throughput than an 8-bit baseline in its own online stress tests. It points to an SGLang-based deployment path for NPUs.

Meituan, a food-delivery and local-services company, is not in the site’s organisation list, so the entry carries no organisation tag. At 1.6 trillion parameters it is among the larger openly licensed models in the record, and the release continues a pattern of Chinese firms outside the main AI labs publishing frontier-scale weights.