LongCat-2.5 Is Live: Meituan’s 1.6T Multimodal Agent Model, API Access and 10 Million Free Tokens

Meituan has moved LongCat-2.5 from preview to launch. On November 2, 2026 its official account announced on X "our latest natively multimodal agentic model with 1.6T total parameters and ~48B active," built "for long-horizon agentic tasks," with "stronger visual understanding and agentic coding" for work from "complex coding projects to visual coding such as creative design and 3D tasks." The API is open now.

We covered the earlier LongCat 2.5 preview in September. This post covers what the launch adds, what the official API docs confirm, and what we could not verify.

LongCat-2.5 at a glance

What does the launch add over the preview?

The September preview was reported through aggregator and gateway listings. It had no official benchmark table. The November 2 launch comes from Meituan's own account and its docs. Meituan's official thread adds details on how the model was trained and what it can do:

• Training. Meituan says LongCat-2.5 "scales an end-to-end agentic RL pipeline across data, environments, algorithms and harnesses for long-horizon tasks," and that its MOPD training retains the previous model's expert groups. The thread adds that per-cut prefixes and segment-capped student rollouts keep long responses stable. We could only see the first lines of that post, so we cannot summarize the rest.
• Demos. The thread shows examples: a real-time fluid simulation and VFX tool that generates interactive web pages, a 3D artifact showcase that renders a bronze sacred tree from the Sanxingdui site with drag-to-rotate and scroll-to-zoom controls, a business operations analysis, and a customer feedback report that produces an interactive HTML page.
• Access. The API is available on the LongCat platform, with credits automatically added to accounts.

The demos are Meituan's own, shown as videos on X. They show what the model can produce on chosen prompts. They are not a test.

What do the API docs confirm?

The LongCat API quick start lists the model as "LongCat-2.5 — High-performance Agentic model" and gives these facts:

• Two API formats. An OpenAI-style endpoint at https://api.longcat.chat/openai supporting chat completions, and an Anthropic-style endpoint at https://api.longcat.chat/anthropic supporting the messages endpoint. The docs show the standard OpenAI and Anthropic SDKs working by changing the base URL.
• Limits. "1M token context window with a maximum output length of 128K tokens."
• Rate limits. A 429 response with a retry_after value when you exceed limits. The docs recommend exponential backoff.
• Tool support. The docs sidebar lists guides for Claude Code, Codex, OpenCode, Cline, Kilo Code, Cherry Studio and Hermes Agent, among others.

A minimal call with the OpenAI SDK, taken from the docs' pattern:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_APP_KEY",
    base_url="https://api.longcat.chat/openai",
)

response = client.chat.completions.create(
    model="LongCat-2.5",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1000,
)
print(response.choices[0].message.content)

The Anthropic-format endpoint means tools built for Claude's message API can point at LongCat with a different base URL and key. That is convenient for testing, but check the tool's own docs, because some features such as tool calls and image inputs may behave differently on a compatible endpoint.

What we could not verify

The comparison point matters. A lab's own benchmark table picks the rivals, the settings and the tasks. Wait for independent runs, such as those on Artificial Analysis, before treating any ranking as settled.

How does it fit Meituan's LongCat line?

Meituan, the Chinese food-delivery and local-services company, has put out a run of models under the LongCat name. Our earlier posts track the line:

1. LongCat-2.0: a 1.6T mixture-of-experts open model trained on AI ASIC superpods, aimed at coding agents.
2. LongCat 2.5 preview: the September listing with image input and a long context.
3. LongCat-2.5 launch (this post): native multimodality, the official docs and the token credit.

The parameter count stayed at 1.6T from 2.0 to 2.5, so the claimed gains come from training (the agentic reinforcement-learning pipeline) and from adding vision, not from a bigger model. Mixture-of-experts models activate a fraction of their parameters per token, which is why 48B active matters for serving cost more than the 1.6T total.

What do we know from the 2.0 release that may carry over?

LongCat-2.0 is the best guide to how Meituan ships. According to our LongCat-2.0 coverage, that model was a 1.6T mixture-of-experts system with about 48B active parameters and a 1M-token context, and its weights and inference code went live under the MIT licence on July 5, 2026. Meituan's own comparison table for 2.0 reported 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench 2.1, and that same table showed Claude Opus 4.8 ahead on both. In other words, the company has published tables where its model does not lead, which is a point in favour of reading its tables as informative rather than pure marketing.

That history gives three things to watch for with 2.5, none of them confirmed:

1. Open weights later. The 2.0 model started with API access and then released weights. Meituan has not said 2.5 will follow, so do not plan around it.
2. A comparison table with named rivals. The 2.0 table named competitors and settings. If a 2.5 table appears, check which model versions it uses, and whether the competitor numbers come from the competitor's own reports or from Meituan's runs.
3. A technical report. The training description in the thread, covering an agentic reinforcement-learning pipeline across data, environments, algorithms and harnesses, is brief. A report would let outsiders judge how much of the gain is from the pipeline.

Questions developers are likely to ask

Is this a drop-in replacement for my current model? Not automatically. The endpoints are compatible, but prompts tuned for another model can behave differently. Treat the swap as a new model evaluation, not a configuration change.

Is 1M context useful in practice? A large window is a ceiling, not a guarantee of quality at that length. Test retrieval and reasoning on documents near your real sizes, and watch cost and latency, because long prompts are slow and billed by tokens once your free credit is used.

What does native multimodal mean? Meituan describes the model as natively multimodal, meaning image understanding is part of the model rather than a separate vision add-on. The docs sidebar lists a "Visual Understanding" section under the API overview, which is where to look for input formats and limits.

Should I put it in my coding agent? The docs list setup guides for several coding tools. That makes a trial easy. For production code, keep your tests and review process in place, as with any model, and compare error rates on your own repositories.

Can I trust the free credit to last? It is a launch offer stated in a social post. Check your balance page and the pricing page before building a workflow on it.

Why it matters for builders

It is another strong non-US option on the API. The 1M context, the Anthropic-style endpoint and the free credit make it cheap to trial. See how other Chinese models compare in our posts on Kimi K3.1 Pro and GLM 5.3 on Bedrock.

Small-model results are worth testing against yours. Teams have found that a specialized or cheaper model can beat frontier models on a narrow task. Coinbase's fine-tuned Qwen3.5 9B on fraud is one case. Run your own evaluation set before switching.

Data location and terms. A hosted API from a Chinese company means your prompts go to its servers. Check the platform's privacy terms, avoid sending regulated or confidential data in a trial, and use a separate key with a spending limit.

How to evaluate it in an afternoon

1. Sign up and create an API key, and note the credits shown in your balance.
2. Point an existing OpenAI-style or Anthropic-style client at the LongCat base URL, changing only the model name.
3. Run 20 to 50 of your own real tasks, including one long-context job and one image task.
4. Record pass rate, latency and tokens used, and compare against your current model on the same tasks.
5. Decide only after reading the pricing page and the data terms.

What to watch next

• Meituan's launch blog and any technical report, with the full benchmark table and settings.
• A Hugging Face or GitHub release, which would tell you whether weights are coming.
• Independent evaluations and price listings from gateways.
• Whether the Anthropic-style endpoint supports the tool-use and caching features your harness needs.

Related reading

• LongCat 2.5 preview: Meituan's 1.6T model for agents
• LongCat-2.0: open-source MoE coding agent
• Artificial Analysis Intelligence Index v5
• Kimi K3.1 Pro: open weights, benchmarks and price
• GLM 5.3 on Amazon Bedrock
• Coinbase's fine-tuned Qwen3.5 9B beats Opus on fraud
• Sources: Meituan LongCat on X, LongCat API docs

Details are accurate as of November 2, 2026. Launch credits and limits can change.