The Pulse
Meituan Releases LongCat-2.5, a 1.6T-Parameter Agent Model
Meituan’s LongCat-2.5 combines visual input with agentic coding and reports gains over its predecessor on several benchmarks. Its published results also trail competitors on some tests, and the company notes that some rival scores come from
AI.info Team · Written with AI · fact-checked ·
LongCat-2.5 is a 1.6-trillion-parameter model with about 48 billion parameters active per token, built for multimodal input and long-running agent tasks. Meituan’s LongCat post describes improvements to visual understanding and agentic coding, while its benchmark table shows a mixed picture: the model beats some listed rivals on certain tests and falls behind others.
The LongCat API change log lists the model as available on November 2, 2026. The company says the system is intended for work such as complex codebase development and full-project builds, as well as visual coding, creative design and 3D tasks.
Visual tokens and long-running agent work
LongCat represents visual signals as multi-level discrete tokens, allowing its backbone to process visual and textual inputs in a shared token-based format. The post says the design combines native-resolution visual tokenization, offline visual preprocessing and visual embeddings learned jointly with the backbone. It is intended to support visual tasks while simplifying computation and workload management during training.
“We introduce LongCat-2.5, a native multimodal model for long-horizon agentic tasks.”
LongCat’s authors write
For agent work, the team says it scaled training across data, environments, algorithms and harnesses. Its coding-task pipeline generates tasks, interactive environments and verifiable tests; it also uses real-world agent failure trajectories to create reproducible training tasks and add data for capability gaps. Training mixes tasks and scaffolds associated with different benchmarks, which the post says improves generalization across setups.
The post also describes long-context training and four-bit weight-and-activation inference on domestic AI accelerators. In online stress tests, LongCat reports 43.3% higher prefill throughput and 42.77% higher decode throughput than its W8A8 INT8 baseline. Those are company-reported results for that comparison.
Benchmarks show gains and gaps
On Terminal Bench v2.1, LongCat-2.5 scores 88.4, compared with 70.8 for LongCat-2.0; Claude Opus 5 scores 89.1 and DeepSeek-V4.1-Flash 90.6. On PaperBench, LongCat-2.5 scores 82.1, above Claude Opus 5’s 79.1 but below GPT-5.6-Sol’s 83.3.
On BrowseComp, LongCat-2.5 scores 91.9, ahead of LongCat-2.0 at 79.9 and all the listed rivals in the table, including kimi-k3 at 91.2 and DeepSeek-V4.1-Flash at 91.5. The picture differs on Vision2Web: LongCat-2.5 scores 78.6, below Claude Opus 5 at 80.2 and kimi-k3 at 79.0.
Meituan says LongCat-2.5 results generally use thinking mode, temperature 1.0 and top-p 0.95. The post marks some competing-model scores as drawn from public reports, blog posts or leaderboards; other rival results are self-run and described as indicative. That means the table combines results from different sources and evaluation conditions rather than a single uniform head-to-head test.
API access and limits
LongCat’s documentation lists support for OpenAI- and Anthropic-format API requests and a context window of up to 1 million tokens, with a maximum output length of 256K tokens. The release notes also list compatibility with Claude Code, Codex and Hermes, among other agent harnesses.