LongCat-2.5

// Overview

LongCat-2.5 is a native multimodal Mixture-of-Experts model from Meituan's LongCat team, built for long-horizon agentic tasks. It has 1.6 trillion total parameters with approximately 48 billion activated per token. The LongCat API Platform changelog lists it as available on November 2, 2026, following a LongCat-2.5-Preview on September 25, 2026; the weights are published on Hugging Face under the MIT license.

The model builds on LongCat-2.0 and combines visual understanding with agentic coding for complex codebase development, full-project builds, creative design and 3D tasks. Following LongCat-Next, native-resolution visual features are converted into discrete tokens and aligned with text in a shared semantic space. On the API it accepts images and videos alongside text, with a 1M-token context window and up to 256K output tokens.

Meituan's model card credits long-sequence training, large-scale agentic reinforcement learning on real-world engineering tasks, and multi-teacher learning (specialised agent, reasoning and interaction teachers distilled with full-vocabulary KL distillation and Tree Attention). It also says training and serving run on domestic AI accelerators: the expert weights and activations are quantized to W4A4 for native INT4 computation, and the published deployment guide targets NPUs through SGLang-FluentLLM.

LongCat-2.5 is a hybrid model with a thinking mode and a non-thinking mode, switched per request with enable_thinking. In Meituan's evaluation table it scores 88.4 on Terminal Bench v2.1, 91.9 on BrowseComp, 83.1 on MCP Atlas and 63.1 on HLE with tools, set against GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash. Meituan marks which competitor cells are cited from public reports; the unmarked scores, including every LongCat-2.5 figure, are in-house measurements it calls indicative only.

// Benchmarks

This model's scores

1. BrowseComp91.9%
2. Terminal Bench v2.188.4%
3. MCP Atlas83.1%
4. PaperBench82.1%
5. Vision2Web78.6%
6. HLE (w/ tools)63.1%
7. DeepSWE (v1.1)64.6%
8. NL2Repo60.5%
9. AutomationBench (v1.0.6)51.8%
10. Terminal Bench v4.038.5%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

LongCat API Platform discounted price, labelled limited-time (¥2 / ¥0.04 / ¥8 per 1M tokens in CNY).

Pricing source ↗

// Strengths

• 1.6T-parameter Mixture-of-Experts with about 48B active per token, released as MIT-licensed open weights
• Native multimodal input: images and videos alongside text
• 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp in Meituan's own evaluation
• 1M-token context and up to 256K output tokens on the LongCat API
• OpenAI- and Anthropic-compatible API endpoints, documented for Claude Code, Codex, Hermes Agent and other harnesses
• Hybrid thinking and non-thinking modes

// Best for

• Reach for it for long-horizon coding agents — full-project builds and repository-scale work — driven from harnesses such as Claude Code or Codex.
• Reach for it for visual coding: turning screenshots, designs or video into working front-end or 3D code.
• Reach for it as a permissively licensed, self-hostable frontier-scale MoE if you have NPU infrastructure; the published deployment guide targets NPUs.
• Look elsewhere for small-footprint local use: at 1.6T total parameters the weights need multi-accelerator serving.

// How to access