LongCat-2.5 Explained: Meituan's 1.6T Agent Model and API

LongCat-2.5 is a Mixture-of-Experts (MoE) model released by China's Meituan in November 2026, with roughly 1.6T total parameters and about 48B active per token. It is natively multimodal, handling text and images in a single token format, and is aimed at coding agents and long-running autonomous tasks. For now it is available only through the chat service and the API. The weights are not released (Meituan has signaled GitHub and Hugging Face releases, but no date or license has been announced). In other words, unlike LongCat-2.0, you cannot yet download it and run it locally.

The API offers both OpenAI-compatible and Anthropic-compatible endpoints, and the official docs include setup guides for Claude Code, Codex, OpenCode, and more. This article sticks to what could be confirmed from the official documentation and press coverage: capabilities, pricing, connection steps, local-run status, and caveats.

Sources disagree on the release date, ranging from November 1 to 3. The official change log entry is dated 2026-11-02, and Meituan's own description points to an announcement on November 1 with API availability on November 2. A preview (LongCat-2.5-Preview) had already launched on September 25.

LongCat-2.0 was announced on June 30, 2026, and its weights were published on Hugging Face under the MIT License in July. It is a text-centric model with the same scale as LongCat-2.5 (1.6T total, about 48B active). The main changes in 2.5 are these three.

- Native multimodality: images are said to be processed in the same token format as text rather than through a separate vision encoder. Meituan emphasizes visual coding: reading screenshots, UI designs, and diagrams and turning them into code
- Stronger long-horizon agents: built for understanding large codebases and building whole projects through multi-step reasoning and tool use. Reports say real-world agent failures were converted into reproducible training tasks for reinforcement learning
- Compatibility with existing agent harnesses: the official change log explicitly names Claude Code, Codex, and Hermes

The benchmark jumps are large too. According to figures attributed to the official page, Terminal Bench v2.1 rose from 70.8 for LongCat-2.0 to 88.4, and BrowseComp from 79.9 to 91.9. All of these are self-reported by Meituan.

- Coding agents: long jobs that include terminal work, such as multi-file edits, running tests, and iterative debugging
- Visual coding: building web screens or reproducing designs from images and screenshots, evaluated on benchmarks such as Vision2Web
- Creative and 3D tasks: the official description also lists creative design and 3D code generation (3DCodeBench)
- Tool use and browsing: Meituan reports strong numbers on evaluations that chain search and tool calls, such as BrowseComp and MCP Atlas

Output is text only; image or audio generation does not appear in the official description. Sources differ on video input, and we could not confirm it officially, so we describe the input as text and images.

LongCat-2.5 is available through the official chat service and the API platform (longcat.chat). When we checked Meituan's official Hugging Face organization (meituan-longcat), it listed LongCat-2.0 and its FP8 and INT8 variants, but not LongCat-2.5. Press reports say GitHub and Hugging Face releases are planned, with no timeline and no confirmation that weights will be included.

- Weight release: only announced as a plan; no date officially stated
- License: not announced for LongCat-2.5. It is not settled that LongCat-2.0's MIT License will carry over
- Commercial terms: API use follows the LongCat API terms of service, which are worth reading before adoption

According to the official docs (Quick Start), there are two endpoint families, and the model name for both is LongCat-2.5.

The docs menu includes Chat Completions, a Responses API, Anthropic Messages, Visual Understanding for images, and model list endpoints. API keys are shown only once at creation, so store them securely. A minimal OpenAI-compatible example follows (put your key in the LONGCAT_API_KEY environment variable).

On sampling, coverage indicates that temperature is fixed at 1.0 with thinking enabled and 0.7 with thinking disabled, with other values overridden. Exceeding rate limits returns HTTP 429 (rate_limit_exceeded, with retry_after), and the docs recommend exponential backoff.

The official docs' Tools section has guides for Claude Code, Codex, and OpenCode, as well as CatPaw, Cline, Kilo Code, Cherry Studio, Chatbox, CodeBuddy, OpenClaw, Hermes Agent, CC Switch, WorkBuddy, and DeepSeek Harness. Because an Anthropic-compatible endpoint exists, Claude Code only needs the base URL and model name swapped through environment variables. An example for ~/.claude/settings.json follows.

This example is based on a developer article describing LongCat's Claude Code guide; variable names and recommended values may change, so check the official guide before configuring. The one thing to watch is context length: Claude Code assumes 200K tokens for unrecognized model IDs and auto-compacts at that size, so using the full 1M tokens of LongCat-2.5 requires setting CLAUDE_CODE_MAX_CONTEXT_TOKENS, as that article points out.

For Codex and OpenCode, you register the OpenAI-compatible endpoint (https://api.longcat.chat/openai) as a custom provider. The exact config format changes with each tool's version, so follow the official Tools page.

One caveat first: the official Quick Start does not list prices, only a link to a separate LongCat-2.5 pricing page, and we could not retrieve that page. The figures below are what multiple third-party reports agree on, and they have not been re-checked against the official price table. Some outlets also reported that no pricing had been announced at launch, so the picture is mixed.

These are described as limited-time discounted prices that may change. LongCat-2.0 launched at $0.30 input and $1.20 output and later moved to its current $0.75 and $2.95, so a similar move to standard pricing for 2.5 is plausible. Our LongCat-2.0 pricing and VRAM article shows how that price evolved. The safe summary is: discount-period rates have been reported, but the confirmed standard price needs checking on the official page.

Not at the moment, because the weights are unreleased. There is no way to run LongCat-2.5 locally; use is limited to the API and the official chat. Even after a release, a 1.6T MoE will still need memory for all of its total parameters. For LongCat-2.0, which has the same 1.6T and about 48B active, even INT4 quantization is estimated to need around 970GB, which is not realistic for a personal PC. See our LongCat-2.0 requirements and VRAM article for details.

LongCat-2.0's weights are already public on Hugging Face under the MIT License, including FP8 and INT8 variants. If you want to experiment locally, LongCat-2.0 is the practical choice for now. We plan to revisit the requirements once LongCat-2.5 weights appear.

All figures below are Meituan-published (self-reported) values with no third-party reproduction. The numbers come from explainer sites that cite the official page; we could not read the primary official blog text directly. Treat them as reference points.

On positioning against other labs, Meituan reportedly says its scores on some mainstream benchmarks fall between Claude Opus 4.8 and Claude Opus 5, and beat Opus 5 on some. The official page, however, is also described as calling the model competitive but not dominant, with Claude Opus 5 ahead on several hard repo-scale benchmarks. Because the presenter chooses the comparison set and conditions, the baseline advice is to test on your own workload.

- Where data goes: the API runs on a Chinese company's infrastructure. Before sending internal code, personal data, or confidential documents, review the terms, data handling, and retention, and compare against your own information-security policy
- Prices may change: they are reported as discounted, and a move to standard pricing is possible. Budget against the official pricing page
- Max output inconsistency: the official docs mention both 128K and 256K. Test the real limit if you need very long outputs
- Claude Code context setting: using 1M tokens needs an extra environment variable, otherwise it may auto-compact at 200K
- Self-reported benchmarks: third-party reproduction is still scarce. Verify performance on your own tasks
- Regulatory movement: a draft Chinese export-related catalogue published October 16 is open for comment until November 15, and Meituan has not said whether LongCat-2.5 is covered. Availability terms could change

- LongCat API Platform Change Log: https://longcat.chat/platform/docs/change-log
- LongCat API Quick Start: https://longcat.chat/platform/docs/
- Hugging Face meituan-longcat: https://huggingface.co/meituan-longcat
- LongCat-2.5 model page (longcatai.org): https://www.longcatai.org/models/longcat-2-5
- DEV Community, Meituan's LongCat-2.5 Drops Into Claude Code: https://dev.to/gabrielanhaia/meituans-longcat-25-drops-into-claude-code-at-030120-its-setup-caps-you-at-200k-4pd0
- Inside AI, Meituan launches LongCat-2.5: https://insideai.news/news/agentic-ai/meituan-longcat-2-5/14286/
- APIYI, LongCat 2.5 Guide: https://help.apiyi.com/en/longcat-2-5-native-multimodal-agent-model-guide-en.html
- LLM Stats, LongCat-2.5: https://llm-stats.com/models/longcat-2.5
- Technology.org, Meituan Releases LongCat-2.5: https://www.technology.org/2026/11/04/meituan-longcat-2-5-multimodal-ai-agent-model-chinese-chips/