Image: Meituan LongCat
Meituan has released LongCat-2.5, a natively multimodal model built for long horizon agentic tasks. The LongCat team announced the model on X on November 2, describing it as having 1.6 trillion total parameters with about 48 billion active per token.
Introducing LongCat-2.5, our latest natively multimodal agentic model with 1.6T total parameters and ~48B active.
— Meituan LongCat (@Meituan_LongCat) November 2, 2026
According to the official LongCat blog, the model builds on LongCat-2.0 with stronger native visual understanding and agentic coding. Meituan says it is aimed at work ranging from complex codebase development and full project builds to visual coding, creative design and 3D tasks. The blog lists GitHub and Hugging Face releases as coming soon, and the LongCat Hugging Face page still lists LongCat-2.0 as its newest large model at the time of this writing. Access is currently through LongCat's chat and API, as AIBase also reported.
The main architectural change is how the model handles images. LongCat-2.5 turns visual signals into multi level discrete tokens, so the backbone processes images and text in the same token based format. Meituan says it applies residual vector quantization to native resolution features from the semantic aligned encoder first used in LongCat-Next. Because the tokenizer is fixed, visual token IDs can be computed and cached ahead of training, which the team says simplifies packing long sequences and balancing load across hardware.
Image: Meituan LongCat
On the training side, Meituan describes a large scale reinforcement learning pipeline for long tasks. It synthesizes coding tasks and test environments from filtered GitHub pull requests and linked issues, checked with fail to pass and pass to pass tests, and also turns real agent failures into repeatable training tasks. To limit reward hacking, the team says it disables network access in task environments, removes git history and cached files, and uses an LLM judge to audit trajectories.
Meituan's published benchmark results put LongCat-2.5 close to leading closed models. The company reports 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0, and 38.5 on Terminal Bench v4.0 against 37.3 for GPT-5.6-Sol. It also lists 91.9 on BrowseComp, 51.8 on AutomationBench, 83.1 on MCP Atlas and 25.5 on Program Bench. On visual tasks Meituan reports 55.4 on 3DCodeBench and 78.6 on Vision2Web.
Image: Meituan LongCat
The same table shows the model trailing rivals in several areas. Claude-Opus-5 scores 53.9 on Terminal Bench v4.0 and 37.0 on Program Bench in Meituan's chart, and GPT-5.6-Sol leads on DeepSWE v1.1 with 72.7 against 64.6 for LongCat-2.5. Meituan notes that some competitor scores were run by its own team and are indicative only, while others come from public reports or leaderboards.
Meituan also highlights inference work on domestic AI accelerators. A W4A4 setup quantizes both expert weights and activations to 4 bits, and the company says it matches the unquantized model on all evaluated downstream tasks. Against a W8A8 INT8 baseline, Meituan reports prefill throughput gains of 43.3% and decode gains of 42.77%, with time to first token down 15.91%.
The release follows LongCat-2.0, which Meituan published with open weights on Hugging Face. Whether LongCat-2.5 weights will follow on the same terms has not been detailed beyond the coming soon notice on the blog.
Comments