AnswersPublished18 sources
Meituan LongCat-2.5: What Changed From LongCat-2.0 and the Preview
Released November 2, 2026, LongCat 2.5 is a multimodal model for sustained agent tasks, reported at 1.6 trillion total and about 48 billion active parameters per token. The main stated upgrades are visual input processed alongside text, training for longer engineering tasks, and support for visual coding, design and...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Released November 2, 2026, LongCat 2.5 is a multimodal model for sustained agent tasks, reported at 1.6 trillion total and about 48 billion active parameters per token.
- The main stated upgrades are visual input processed alongside text, training for longer engineering tasks, and support for visual coding, design and 3D work.
- Meituan reports 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat 2.0; that result alone does not establish broad superiority over other models.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Meituan’s LongCat-2.5, released on November 2, 2026, and how does it build on LongCat-2.0 and the September 25 API preview in terms. Article summary: LongCat-2.5 is Meituan’s native multimodal, mixture-of-experts model aimed at sustained agent work, particularly complex coding projects, full-project engineering, and visually guided development; launch reporting descri. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fak
Meituan released LongCat-2.5 on November 2, 2026, as a native multimodal model for long-running agent tasks. Building on LongCat-2.0 and the September 25 LongCat-2.5-Preview, it is presented as a step toward agents that can carry complex engineering work through multiple stages—not just generate isolated code.6
32
What LongCat-2.5 is designed to do
Meituan describes LongCat-2.5 as a 1.6-trillion-parameter model with roughly 48 billion parameters active per token. These are reported figures; one supplied source instead lists 480 billion active parameters, so the active-parameter count is not consistent across all reporting.8
32
1
The release’s stated focus is end-to-end work in complex codebases and building complete projects. Meituan says it added more long-horizon task trajectories to long-sequence training and used large-scale reinforcement learning to improve agentic coding. Its training pipeline also includes data synthesis and environment construction across agent harnesses, and turns real agent failures into reproducible, verifiable training tasks.32
How its multimodal approach works
LongCat-2.5 represents visual signals as multi-level discrete tokens and processes visual and text inputs in a shared model backbone, according to launch coverage. The aim is to connect visual understanding with coding, including visual coding and tasks involving creative design and 3D.32
35
Meituan also says it optimized long-context training and W4A4 inference on domestic computing infrastructure. The supplied reporting does not provide independently measured efficiency gains, so these should be read as descriptions of the company’s approach rather than proof of a particular performance advantage.32
Reported benchmarks: gains over LongCat-2.0, with limits
The available benchmark figures show gains on several listed tests, but they are reported results—not independent verification. Meituan’s model page lists the following scores for LongCat-2.5 and LongCat-2.0:38
| Benchmark | LongCat-2.5 | LongCat-2.0 |
|---|---|---|
| Terminal Bench v2.1 | 88.4 | 70.8 |
| AutomationBench | 51.8 | 15.8 |
| BrowseComp | 91.9 | 79.9 |
| PaperBench | 82.1 | 45.2 |
| Program Bench (Almost Resolved) | 25.5 | 2.5 |
A separate report describes the benchmark scores as vendor-reported and not independently verified.6 Results from individual tests do not establish that a model is best across all tasks or comparisons. They are most useful as an indication of the areas Meituan is targeting and the improvements it reports against its own earlier model.
Release, API and agent-tool support
The September 25 preview preceded the November 2 full release, according to secondary coverage.6 LongCat’s platform changelog now lists LongCat-2.5 as available and describes support for visual tasks and complex engineering workflows.
35
The LongCat API supports OpenAI-compatible and Anthropic-compatible formats.18 The platform also names Claude Code, Codex and Hermes among compatible agent harnesses, while its OpenCode guide explains how to authenticate and select LongCat as a provider.
35
36
The takeaway
LongCat-2.5’s clearest stated distinction is its combination of long-horizon agent training and native visual input, aimed at carrying coding and related creative tasks through more of the workflow. The model’s reported scale and benchmark results provide useful context, but they should be treated as reported claims—not as independent confirmation of general-purpose leadership.6
32
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Meituan LongCat-2.5: What Changed From LongCat-2.0 and the Preview"?
Released November 2, 2026, LongCat 2.5 is a multimodal model for sustained agent tasks, reported at 1.6 trillion total and about 48 billion active parameters per token.
What are the key points to validate first?
Released November 2, 2026, LongCat 2.5 is a multimodal model for sustained agent tasks, reported at 1.6 trillion total and about 48 billion active parameters per token. The main stated upgrades are visual input processed alongside text, training for longer engineering tasks, and support for visual coding, design and 3D work.
What should I do next in practice?
Meituan reports 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat 2.0; that result alone does not establish broad superiority over other models.
Sources
dev.toMeituan's LongCat-2.5 Wants Your Agent to Finish the Whole Job — Terminals, Browsers, Spreadsheets and All
ithome.com美团龙猫发布LongCat-2.5 原生多模态AI 模型：1.6T 参数
technode.comMeituan launches LongCat-2.5 multimodal AI model for extended coding tasks · TechNode
ai-all.infoLongCat-2.5 – In-Depth Evaluation of Meituan's Native ...