AI Tools

LongCat-2.5 Open Model: Price, Features and Use in India

Meituan has released LongCat-2.5, a 1.6-trillion-parameter open-weight agent model under the MIT licence. Here is what it does, what the API costs and how Indian teams can use it.

Developer typing on a laptop at a desk, illustrating the LongCat 2.5 open AI model for coding agents

LongCat-2.5 is a large open-weight AI model from Meituan’s LongCat team, built for AI agents that code, browse and use tools for long periods. Meituan launched it on 2 November 2026 and released the weights under the MIT licence on 15 November 2026. Downloading the weights is free; the API is paid.

Most Indian teams use AI through ChatGPT, Gemini or Claude. LongCat-2.5 matters because it is a near top-tier agent model that anyone can download and use commercially, and its API price is reported to be far lower than leading closed models. For Indian developers, agencies and SaaS startups, that can change the cost of building AI features. For non-technical owners, our AI for small business in India guide is a better starting point.

Key takeaways

  • LongCat-2.5 was launched by Meituan on 2 November 2026, according to TechNode.
  • The weights went public on Hugging Face under the MIT licence on 15 November 2026, as reported by The Decoder.
  • It is a mixture-of-experts model with about 1.6 trillion total parameters and roughly 48 billion active per token.
  • Its API supports OpenAI and Anthropic request formats, so it can plug into tools like Claude Code.
  • Launch API pricing is reported as limited-time and low; official India availability is not confirmed.

At a glance

Product Made by Launched Price Available in India
LongCat-2.5 (open-weight AI model) Meituan (LongCat team) 2 November 2026 (weights: 15 November 2026) Weights free (MIT); API about US$0.30 input / US$1.20 output per million tokens (reported, limited-time) Not confirmed (weights downloadable worldwide)

What is LongCat-2.5?

LongCat-2.5 is a large language model made by the LongCat team at Meituan, the Chinese company best known for food delivery. The model card on Hugging Face describes it as a native multimodal model for long-horizon agent tasks.

In simple words, it is designed to work on a job for a long time: read a big codebase, run commands, browse the web, call tools and keep going until the task is done. It accepts text and images and answers in text.

According to TechNode’s report on 2 November 2026, Meituan first made the model available through its own chat service and API platform. The open weights followed about two weeks later.

The Decoder reported on 15 November 2026 that LongCat-2.5 was trained entirely on Chinese hardware, and Meituan’s sample self-hosting setup uses 194 Huawei Ascend chips rather than Nvidia GPUs. That is why the launch drew attention during the week of Nvidia’s leadership crisis, which we covered in what Jensen Huang’s exit means for India’s AI plans.

Key features

  • Open weights under MIT licence: businesses can download, modify and use the model commercially, according to the Hugging Face model card. Training data and full training code are not part of the release.
  • Mixture-of-experts design: about 1.6 trillion total parameters, with roughly 48 billion used for each token, which keeps running costs lower than a dense model of the same size.
  • Long context: The Decoder reports up to 1 million input tokens and up to 128,000 output tokens through the API. Other reports give different output limits, so treat this as not fully confirmed.
  • Thinking on or off: the model card shows a hybrid mode, where reasoning can be switched on for hard tasks and off for quick replies.
  • Tool calling: the model supports function and tool calls, which is the base for building AI agents.
  • OpenAI and Anthropic compatible API: existing code written for those formats can be pointed at LongCat with small changes.
  • 4-bit version: Meituan also published a W4A4 quantised version, which it says runs about 43% faster at prompt processing than an 8-bit version.

How good is LongCat-2.5 compared to Claude and GPT?

On Meituan’s own tests, LongCat-2.5 is close to the best closed models on several agent benchmarks, but not ahead on all of them. These scores are company-reported and not independently verified.

Benchmark (Meituan’s figures) LongCat-2.5 Comparison
Terminal Bench v2.1 88.4 Claude Opus 5: 89.1, GPT-5.6 Sol: 88.8, DeepSeek-V4.1-Flash: 90.6
Terminal Bench v4.0 38.5 Claude Opus 5: 53.9, GPT-5.6 Sol: 37.3
DeepSWE 64.6 LongCat-2.0: 9.7, DeepSeek-V4.1-Flash: 74.2
BrowseComp 91.9 Highest in Meituan’s table (run in Meituan’s own setup)

The practical reading: it is a strong agent and browsing model, a big jump over LongCat-2.0, but Claude Opus 5 still leads on the hardest new coding test. If you want to understand what “agent” means here, read our simple guide to AI agents.

Price and availability in India

The weights are free to download from Hugging Face. Using the hosted API costs money, and the official price list should be checked before you build on it.

Plan Price Best for
Open weights (Hugging Face) Free download, MIT licence Large companies and labs with serious GPU or NPU servers
LongCat API (launch offer) About US$0.30 per million input tokens and US$1.20 per million output tokens, as reported by several outlets; The Decoder lists about €0.27 and €1.07. Marked limited-time. Developers, agencies and startups building AI features
LongCat chat website (longcat.ai) Free usage limits not confirmed Trying the model before paying

In rupees, the reported launch rate works out to roughly ₹26 per million input tokens and roughly ₹100 per million output tokens. This is approximate, at about ₹85 to a US dollar, and the offer can end at any time.

Is LongCat-2.5 available in India?

The open weights can be downloaded from India like any Hugging Face model. Meituan has not published an India-specific statement for its chat app or API, so official API access, payment methods and support for Indian users are not confirmed.

Can a small business run LongCat-2.5 on its own computer?

No. Meituan’s sample deployment uses a cluster of 194 Huawei Ascend accelerators, and the weights are several terabytes. For most Indian businesses, the realistic route is the API or a third-party host, not self-hosting.

How Indian businesses can use it

The use cases below are Cracckk’s explanation of where the model fits, based on Meituan’s published features. They are not test results.

Software agencies and freelancers

Indian dev agencies often bill fixed prices, so AI cost per project matters. Because the API speaks the Anthropic format, a team could point an agent tool like Claude Code at LongCat for routine work such as writing tests or fixing small bugs, and keep a premium model for hard problems. Our Claude Code automation guide explains that workflow.

A simple example: if an agent uses 20 million input tokens and 2 million output tokens in a month, the reported launch rate would be about US$8.40, or roughly ₹715. Real usage varies a lot.

SaaS startups building for India

A startup making a GST helper, a CRM or a WhatsApp bot can test LongCat as a lower-cost model for background jobs like summarising long documents or classifying leads. The 1 million token context, if it holds in practice, suits long invoices, contracts and chat histories.

Research teams and engineering colleges

Because the weights are MIT licensed, institutes with large compute can study and fine-tune the model without licence fees. Students can also read the model card and benchmark notes to learn how agent models are evaluated.

Is LongCat-2.5 safe for Indian business data?

It depends on how you use it. Self-hosted weights keep data on your own servers, but the hosted API sends your prompts to Meituan’s platform, and its data-storage location and retention terms for Indian users are not confirmed.

Before sending customer names, phone numbers or invoices to any AI API, check your obligations under India’s data protection law. Our guides on AI and the DPDP Act and whether AI is safe for business data list the questions to ask. Many companies also have internal rules about using servers based in China, so check those first.

Limitations to know

  • All benchmark numbers come from Meituan and are not independently verified.
  • The API price is a limited-time launch offer with no announced end date.
  • Self-hosting needs very large hardware, and the official guide targets Huawei Ascend chips, not Nvidia GPUs.
  • Output length limits differ between reports.
  • No India-specific pricing, payment or support information has been announced.

How to try it

  1. Open the official LongCat chat website at longcat.ai to try the model in a browser.
  2. Read the LongCat-2.5 model card on Hugging Face for features, benchmarks and the licence.
  3. For API use, sign in to the LongCat API platform, create an API key and check the current price list.
  4. Use the OpenAI-compatible or Anthropic-compatible endpoint in your existing code, and start with small, non-sensitive test tasks.
  5. Compare cost and quality against your current model for one week before switching any production work.

Try this prompt

Use this with LongCat-2.5 or any AI model to plan a small automation for an Indian shop.

You are a software assistant for a small Indian retail business.
We sell clothing on WhatsApp and Instagram and keep orders in a Google Sheet
with columns: Date, Customer Name, Item, Size, Price (INR), Paid (Yes/No).
1. Suggest a simple daily report that shows total sales, unpaid orders and top items.
2. Write the steps to build it with a free tool, in plain English.
3. List what customer data should NOT be sent to an AI tool, keeping India's DPDP Act in mind.
Keep the answer under 300 words and use bullet points.

More ready-to-copy prompts: free AI prompt library.

Alternatives

  • Claude Opus 5 / Opus 5.5 (Anthropic): stronger on the hardest coding tests in Meituan’s own table, but closed and more expensive.
  • GPT-5.6 Sol (OpenAI): a closed model with similar agent scores in Meituan’s comparison and wide availability in India.
  • DeepSeek-V4.1-Flash: another Chinese model that scores higher than LongCat-2.5 on some coding benchmarks in the same table.

FAQ

Is LongCat-2.5 free to use?

The LongCat-2.5 weights are free to download under the MIT licence, which also allows commercial use. However, running such a large model needs costly hardware. The hosted API is paid, with a reported limited-time rate of about US$0.30 per million input tokens. Free usage limits on the chat website are not confirmed.

Is LongCat-2.5 available in India?

Indian users can download the open weights from Hugging Face like anyone else. Meituan has not announced India-specific availability, pricing in rupees or payment support for its chat service and API. So official API access for Indian users is not confirmed, and businesses should check sign-up and payment options themselves.

Who made LongCat-2.5?

LongCat-2.5 was made by the LongCat team at Meituan, a large Chinese technology company known for its food delivery app. Meituan launched the model on 2 November 2026 and published the open weights on Hugging Face on 15 November 2026. The Decoder reports it was trained entirely on Chinese hardware.

Is LongCat-2.5 better than Claude?

Not overall. In Meituan’s own benchmark table, LongCat-2.5 scores close to Claude Opus 5 on Terminal Bench v2.1 (88.4 vs 89.1) but well behind it on Terminal Bench v4.0 (38.5 vs 53.9). Its main advantages are open weights and a much lower reported API price, not top quality.

Can I use LongCat-2.5 with Claude Code?

Reports and LongCat’s API documentation say the API offers an Anthropic-compatible endpoint, so tools like Claude Code can be pointed at it using a LongCat API key. Tool-calling behaviour may differ from Claude models, so test on small tasks first and never use your Anthropic key on another provider.

Sources

Featured image: Laptop Code by Matthew Henry, CC CC0 (via Openverse)

Written by

Cracckk

The Cracckk editorial team explains new AI tools and AI news for Indian small businesses, creators and students in simple English. Articles are researched from official announcements and reputable sources, and each one lists its sources. Spotted an error? Email support@cracckk.com.

Website / profile →

Leave a Reply

Your email address will not be published. Required fields are marked *