Meituan releases LongCat-2.5: 1.6T-parameter multimodal model, 88.4 on Terminal Bench v2.1
Meituan's LongCat team released LongCat-2.5 on November 2, a native multimodal model with 1.6 trillion total parameters and about 48 billion active per token. Meituan claims 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0. The figures are unverified by third parties.
Meituan's LongCat team released LongCat-2.5 on November 2, a native multimodal model with 1.6 trillion total parameters and about 48 billion activated per token, Pandaily reported, citing IT Home. The model is built for long-horizon agent tasks and builds on LongCat-2.0, with upgrades to native visual understanding and agentic coding.
The official release follows LongCat-2.5-Preview, which went live on the LongCat API platform on September 25.
Key specifications
- Total parameters: 1.6 trillion
- Active parameters: about 48 billion per token
- Inference format: W4A4 (4-bit weights, 4-bit activations), run on domestic chips according to Meituan
- Context window: not yet disclosed
- Pricing: not yet disclosed
Training approach
Meituan says it added more long-horizon trajectory data to long-sequence training and applied large-scale long-horizon reinforcement learning. The stated goal is better end-to-end engineering work, such as developing inside complex codebases and building complete projects.
The team says it continues to expand data synthesis, environment construction and mixed training across different agent harnesses. It also converts failure trajectories from real agents into reproducible, verifiable training tasks, so it can add data where the model is weak.
On vision, LongCat-2.5 represents visual signals as multi-level discrete tokens, so the backbone processes images and text in the same token form. Meituan says this native vision, combined with agentic coding, makes the model strong at visual coding and lets it handle creative design and 3D tasks.
Meituan says it optimized long-context training and W4A4 inference on domestic chips. It also cites work on numerical consistency between training and inference, plus long-context rollout optimization, to keep large-scale long-horizon reinforcement learning running efficiently.
Benchmarks (Meituan's figures, not independently verified)
| Benchmark | LongCat-2.5 | LongCat-2.0 |
|---|---|---|
| Terminal Bench v2.1 | 88.4 | 70.8 |
| AutomationBench | 51.8 | 15.8 |
Meituan also lists 91.9 on BrowseComp, 82.1 on PaperBench, 55.4 on 3DCodeBench and 78.6 on Vision2Web. Its charts compare the model with frontier models from other labs, and it does not lead every test. On Program Bench it scores 25.5, against 37 for the top entry.
Availability and tooling
LongCat-2.5 is available at longcat.chat and through the LongCat API. Meituan says GitHub and Hugging Face releases are coming soon.
The LongCat change log says the model works with Claude Code, Codex, Hermes and other mainstream harnesses. Platform documentation includes setup guides for OpenCode, Cline, Kilo Code, CodeBuddy and WorkBuddy.
What this means
The release centers on agentic work rather than single-turn chat. Meituan's own tables put the largest gains over LongCat-2.0 on agent benchmarks: Terminal Bench v2.1 rose from 70.8 to 88.4, and AutomationBench from 15.8 to 51.8. These are Meituan's measurements, and no independent results have been published.
The Program Bench result (25.5 against a top score of 37) shows the model does not lead every comparison Meituan chose to publish. Developers evaluating it for coding agents should test it on their own repositories and harnesses.
Compatibility with Claude Code, Codex and Hermes, along with documented setups for OpenCode, Cline and Kilo Code, means teams can try it inside existing tooling without switching workflows. Context window and pricing are not yet disclosed, which limits cost and long-context planning for now. The announced GitHub and Hugging Face releases have no date yet.
Related Articles
Nace AI open-sources Drex Max v1, a 27B multimodal decision model for Apple Silicon
Nace AI has open-sourced Drex Max v1, a 27B multimodal decision model. MarkTechPost reports it scores 67.52 and runs locally on Apple Silicon.
SenseTime open-sources 9B SenseNova-Loop-9B-Preview, a looped model, under Apache 2.0
SenseTime open-sourced SenseNova-Loop-9B-Preview on October 30 under Apache 2.0. The multimodal model stores 32 decoder layers but runs 56 layer executions per forward pass, and SenseTime claims it matches models of around 27 billion parameters.
Articul8 launches Bivia, an API decision model for enterprise AI agents across text, image, audio and video
Articul8 has introduced Bivia, a decision-making model that lets enterprise AI agents make structured decisions without calling a generative model at every step. It is available by API as a standalone product or inside Articul8's enterprise AI platform. Articul8 reports roughly 70% workflow completion in an internal evaluation, against roughly 30% for the alternative tested.
Liquid AI releases d1-2B, a 1.92B decision model that answers in one forward pass with no generated tokens
Liquid AI has published d1-2B, a 1.92B-parameter decision model built on LFM3-VL-2B with a 128K-token context. It returns typed, calibrated answers in a single forward pass without generating tokens, and Liquid AI claims 5.7 ms per decision on an NVIDIA H100.
Comments
Loading...