LongCat
Meituan · Coding · Other · China AI
LongCat is Meituan's family of open foundation models built for long-context reasoning and agentic coding. It covers fast chat models, deep-thinking variants, and multimodal versions. You can reach all of them through the LongCat chat app or an API that speaks both OpenAI and Anthropic formats. The newest version, LongCat-2.0, packs 1.6 trillion parameters. It also holds a 1M-token context window. That combination is what makes this long-context AI model useful for reading whole code repositories or long documents in one pass.
About LongCat
What Is LongCat
The LongCat foundation model line is Meituan's umbrella name for a set of large language models it develops and mostly open-sources. The line started in September 2025 with LongCat-Flash-Chat. It grew to include a thinking model, an omni model that handles audio and video, image and video generators, and the trillion-parameter LongCat-2.0 that arrived in June 2026. For a regular user, the easiest door in is the LongCat chat app and the longcat.chat website. Both speak plain language. Neither needs any setup.
What separates Meituan LongCat from a generic chatbot is the mix of speed and context length. The earlier Flash models activate only a slice of their parameters per token. Replies come back fast, even on modest hardware. The newer 2.0 model pushes context to a million tokens. Meituan also trained and ran the whole thing on domestic Chinese chips rather than imported GPUs. That last point matters mostly to people who care about supply chains. It's also the reason the pricing stays low.
The honest downside is that you're working with a Chinese-developed model, and some official pages and announcements land in Chinese first. Documentation in English exists for the API, but community help is thinner than what you'd find around GPT or Claude. LongCat also isn't a full SOTA leader across the board. Meituan's own numbers put it ahead of GPT-5.5 and Claude Opus 4.6 on coding benchmarks like SWE-bench Pro, but behind them on broad knowledge and general reasoning tests.
Getting Started
- Go to longcat.chat and sign in with a Meituan scan or a phone number.
- Open the chat box and pick a mode: a light setting for short questions, or deep thinking for longer jobs.
- Paste your task, whether that's a block of code, a long document, or a multi-step instruction.
- For programmatic use, open the API Keys page and copy the default key.
- Point your OpenAI or Anthropic SDK at the LongCat endpoint, swap in the key, and start calling it.
Product Information
A quick look at LongCat's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Developers who want a cheap coding model they can drop into an existing OpenAI or Anthropic workflow without rewriting code.
- Students and tinkerers testing long-context tasks. The free daily token allowance covers light use without a card.
- Teams handling large codebases or long documents who need a model that won't lose track of the middle of a 1M-token input.
Tasks
- Repository-level refactoring
- Document analysis
- Automated agent runs
- Rapid prototyping
Scenarios
- Overnight batch jobs where cost per token matters more than raw brilliance.
- A solo developer who's hit rate limits or bans on another provider and needs a stable fallback that won't disappear mid-project.
- Research setups that need to run inference on domestic hardware rather than imported accelerators.
Key features
Long-Context Reasoning at 1M Tokens
The headline number is a one-million-token context window on LongCat-2.0. Why does that matter? Because it's enough to hold an entire code repository or a book-length document in a single request. You can ask questions that span the whole input instead of stitching together chunks. Meituan built a sparse attention mechanism (LSA) to keep the compute cost linear as the input grows, which is what makes the long window practical rather than just theoretical.
Fast Answers From a Sparse Model
LongCat-Flash models use a mixture-of-experts design in which only a fraction of the total parameters fire for any given token, which keeps compute and cost down on easy inputs. The 1.6T-parameter 2.0 model averages around 48B active parameters per token. It flexes between roughly 33B and 56B based on how hard the input is. Simple questions burn less compute. Complex ones get more. The practical result is fast responses and a lower bill than a dense model of the same size would demand.
OpenAI and Anthropic API Compatibility
You don't have to learn a new SDK. LongCat exposes two endpoint formats, one that mirrors OpenAI's chat completions and one that mimics Anthropic's. If you already have code running against either provider, switching usually means changing the base URL and the model name. That lowers the cost of trying it. It also makes LongCat a realistic drop-in for tools built around those ecosystems.
Open-Weight Models
Most of the LongCat family ships under open licenses, with weights on Hugging Face and GitHub, so you can download them and run them yourself if you'd rather not send data to an API. That range covers chat, thinking, omni, image, and video models. Self-hosting a trillion-parameter model isn't realistic for most people. The smaller variants are within reach of a decent workstation if you want to try them locally.
Agentic Coding Focus
Meituan tuned LongCat-2.0 specifically for coding agents and tool use, not just plain chat. It handles multi-file problems and plans short sequences of tool calls. It also holds up when a task runs for many steps. Meituan reports a SWE-bench Pro score that edges past GPT-5.5 and Claude Opus 4.6, and as an agentic coding model it slots into frameworks like Claude Code and OpenClaw with minimal fuss.
Multimodal and Media Models
Beyond text, the lineup includes LongCat-Flash-Omni for real-time audio and video interaction, LongCat-Image for image generation, and LongCat-Video for text-to-video and video continuation. These all live alongside the chat models inside the same family, so you're not juggling five separate vendors to get text plus media in one place.
Pros and cons
Pros
- A 1M-token context window makes whole-repository and long-document tasks practical.
- Free daily token allowance lets you test real workloads before paying.
- Dual OpenAI and Anthropic API formats cut migration work to near zero.
- Open weights for most models mean you can self-host or fine-tune.
- Low reported pricing relative to comparable Western flagship models.
Cons
- English documentation and community support lag behind the big Western providers, so troubleshooting takes longer.
- General knowledge and reasoning benchmarks trail top-tier models, so it isn't a universal upgrade.
- The strongest model is huge; self-hosting it needs far more hardware than most individuals have.
- Free-tier throughput and rate limits are modest, which can slow down heavy experimentation.
Frequently asked questions
LongCat is a general-purpose AI model family, with extra emphasis on two things: working with very long inputs and acting as the engine behind coding agents. In practice that means summarizing long documents, answering questions about large codebases, and powering automated tool-calling workflows that otherwise stall on shorter-context models.
Related content
Explore related tools, skills, and articles for LongCat.
LongCat Alternatives
Base Code
Base44 (Wix) · Coding · ProductivityBase Code is a shared cloud development environment from Base44. Teammates describe changes to an existing GitHub codebase in chat, preview them live and send them back as pull requests.
Google EmbeddingGemma 2
Google DeepMind · Writing · Image · CodingGoogle EmbeddingGemma 2 is a 740M-parameter open model that turns text, code, images, audio, and video frames into vectors in one 768-dimension space, running on phones and laptops.
GitGlow
GitGlow · Coding · ProductivityGitGlow is a local-first desktop Git client that lets independent developers review their own code and their AI agents' changes in one place before shipping.