Overview
LongCat-2.5-Preview is the new-generation large model preview that Meituan launched on its LongCat API open platform on 2026/9/25, opening API access and a web experience entry at the same time.
It continues the LongCat family's MoE architecture, with roughly 1.6 trillion total parameters and about 48B activated per inference step, natively supporting a 1M-token context window suited to ultra-long documents, codebases, logs, and multi-turn complex tasks.
The biggest change in this iteration is the addition of native image understanding: it can parse image content and support cross-modal Q&A, content summarization, and complex visual reasoning. Coding performance is also further improved, with deep adaptation to Claude Code and other developer environments, efficient collaboration with Claude Code, Hermes, OpenClaw, OpenCode, and Kilo Code, and compatibility with both the OpenAI and Anthropic APIs.
Key Features
- New Native Image Understanding: Parses image content and supports cross-modal Q&A, content summarization, and complex visual reasoning
- Trillion-Parameter MoE Continuation: Roughly 1.6T total parameters with about 48B activated per inference, continuing the domestic trillion-level model lineage
- 1M-Token Context: Native 1M-token window suited to ultra-long documents, codebases, logs, and multi-turn complex tasks
- Deep Coding Agent Adaptation: Further optimized for code generation, understanding, and automated programming, deeply adapted to Claude Code and other developer environments
- Dual API Compatibility: Compatible with both OpenAI and Anthropic APIs, lowering integration and migration cost
- Limited-Time API Pricing: $0.30 input / $1.20 output / $0.006 cache per million tokens, with 5 million free tokens for existing users
Use Cases
- Image understanding and cross-modal Q&A
- Ultra-long document, codebase, and log analysis
- Code generation and automated programming
- Coding agents based on Claude Code and other developer environments
- Multi-turn complex task handling
Pros
- Adds native image understanding, completing multimodal capability
- Continues the trillion-parameter MoE architecture on a stable domestic model lineage
- 1M-token context handles long documents and codebases well
- Deeply adapted to Claude Code and other mainstream developer environments with a mature ecosystem
- Compatible with both OpenAI and Anthropic APIs, low migration cost
- Competitive limited-time pricing with a friendly token grant for existing users
Pricing
Limited-time API pricing (per million tokens): $0.30 input, $1.20 output, $0.006 cache; RMB billing is ¥2 / ¥0.04 / ¥8. Existing users receive 5 million free tokens. A web experience entry is open at the same time, with official pricing and open-source plans to be announced.
Summary
LongCat-2.5-Preview is the next iteration of Meituan's LongCat family, and its biggest highlight is native image understanding, adding multimodal capability to a trillion-parameter MoE model focused on coding agents. Building on the 1.6T / 48B MoE architecture and a 1M-token context, plus deep adaptation to Claude Code and dual OpenAI / Anthropic API compatibility, it keeps developer integration cost low. Meituan has not yet published benchmark results, and official pricing and open-source plans await further disclosure.
Version History
- Meituan launches LongCat-2.5-Preview with native image understanding (2026-09-25): The LongCat API open platform launched LongCat-2.5-Preview, continuing the MoE architecture (about 1.6T parameters, 48B activated) with a native 1M-token context; added image understanding (cross-modal Q&A, content summarization, complex visual reasoning) and code capabilities deeply adapted to Claude Code and other developer environments, compatible with both OpenAI and Anthropic APIs; limited-time pricing of $0.30 input / $1.20 output / $0.006 cache per million tokens, 5 million free tokens for existing users, with no benchmark published yet.