LongCat 2.5: What Meituan's Preview Model Is, and What It Is Not

Ask for LongCat 2.5 on Meituan's own platform and the model you get is called LongCat-2.5-Preview — and that suffix is not decoration. It is the model ID, the string in every code sample, the name on the price list, and the only LongCat 2.5 entry the vendor's model-listing endpoint returns. Meituan's LongCat API Platform put it live under a dated changelog entry on 25 September 2026, alongside the platform's only other model, LongCat-2.0. Nearly three weeks later it has not been renamed, its price has not stopped being labelled promotional, and no one — Meituan included — has published a single benchmark number for it.

That combination is the whole reason this page exists. The interface is fully documented: an OpenAI-compatible endpoint, an Anthropic-compatible one, a published rate card, a published context limit. The capability profile is not documented at all, and one headline feature is described two different ways by two pages of the same documentation set. If you are deciding whether to build on it, the useful thing is a precise list of what is known, what is claimed, and what is simply absent.

What LongCat 2.5 is, and who publishes it

LongCat is Meituan's model line, and the LongCat API Platform is Meituan's own serving surface for it — not a third-party gateway. The platform's FAQ describes its service as "AI model proxy services specifically for the LongCat series models", maintaining compatibility with the Open​AI and Anthrop​ic API formats so that existing SDKs work unchanged. The support address on the vendor's own pages is longcat-team@meituan.com, and the documentation site, the console and the API host all sit under Meituan-controlled domains and Meituan's static-asset CDN.

The catalogue is short. The platform's supported-models table lists exactly two entries, LongCat-2.5-Preview and LongCat-2.0, both described identically as a "High-performance Agentic model". The GET /openai/v1/models endpoint returns those same two and nothing else, each with owned_by set to "LongCat". So "LongCat 2.5" is not a family with variants you have to choose between — there is one 2.5 model, it is the newer of the two on the platform, and it is the one the platform's own pay-as-you-go pricing page leads with.

What the vendor says it is for, in the release note, is three things: image understanding, coding, and compatibility with mainstream development environments — Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code are named. That release note is the vendor's own dated entry, and it is the only place any of those claims appear.

"Preview" is the model ID, not a marketing word

The distinction matters more than it usually does, because on this platform "preview" is load-bearing in three concrete ways.

First, it is the string you must send. Every sample request in the documentation — cURL, Python, the Open​AI SDK example, the Anthrop​ic SDK example — uses LongCat-2.5-Preview, and the Retrieve Model reference uses it as the path parameter. There is no GA alias to send instead.

Second, the price is explicitly provisional. The pricing page puts the words "limited-time" in the column header above every figure, and closes with a sentence worth reading twice: "Prices are subject to change. The prices shown on the LongCat API Platform and your billing records are authoritative." A rate card that tells you it is not the authority for its own numbers is telling you not to build a unit-economics model on it.

Third, no stability commitment exists anywhere. There is no published deprecation window, no versioning promise, no statement about whether the model will keep the same ID when it graduates or whether -Preview will simply be retired in favour of a new name. That is the normal risk of a preview, and the honest way to hold it is: you can call it today, and you cannot assume the call you write today is the call you keep.

What preview does not mean here is restricted access. The platform is self-serve: register, create a key, top up a prepaid balance. Users in Mainland China have to complete identity verification before they can recharge; users outside it can register with a phone number or an email address. Nothing in the documentation gates the model behind an approval process.

Parameters and architecture: the number everyone quotes belongs to LongCat 2.0

Meituan publishes no parameter count for LongCat-2.5-Preview. It is not in the changelog, not on the pricing page, not in the API reference, not in the FAQ. If you have seen "roughly 1.6 trillion total parameters with about 48 billion active" attached to LongCat 2.5, you have seen a LongCat 2.0 figure wearing a 2.5 label — and the giveaway is that the sentence is nearly always qualified with "approximately", because it is being carried across from the older model card rather than read off this one.

The source of that number is Meituan's own Hugging Face model card for longcat-2.0, which opens: "We introduce LongCat-2.0, a large-scale MoE language model with 1.6 trillion total parameters and ~48 billion activated per token." That card goes on to describe LongCat Sparse Attention, pretraining across "hundreds of billions of tokens of 1M-context data", and 135 billion N-gram Embedding parameters included in the model. It is a well-documented architecture, and it belongs to LongCat 2.0. Whether 2.5 inherits that shape is a reasonable expectation and an undocumented fact. This page will not print it as a LongCat 2.5 spec, and neither should anything else you read this month.

Open weights: no

Meituan publishes a great deal on Hugging Face — the meituan-longcat organisation holds around thirty repositories spanning text, image, video and audio models, and the LongCat-2.0 family is there in full: the base model plus FP8 and INT8 siblings, all under MIT. There is no LongCat-2.5 repository. The newest text-generation release in the organisation is LongCat-2.0 from early July 2026, and a search of the Hub for LongCat-2.5 returns no repository from Meituan or from anyone else.

The pattern is telling in the other direction, too. LongCat 2.0's weights being open is visible in the ecosystem that forms around open weights — community 3-bit and 2-bit quantisations, MLX conversions, mixed-precision AutoRound builds. None of that exists for 2.5, because there are no weights for anyone to quantise. Until Meituan publishes a repository, treat LongCat-2.5-Preview as a closed, API-only model, whatever the future of the LongCat line looks like.

The envelope the vendor actually publishes

These are read from Meituan's own documentation and pricing pages on 13 October 2026. Every one of them is the vendor's own figure; none is independently verified, and the last one is not a figure at all.

• Context window — 1,048,576 tokens. The Retrieve Model reference's example response for LongCat-2.5-Preview returns context_length: 1048576, and the quick-start states the same limit in prose as "1M token context window".

• Maximum output — 131,072 tokens. The quick-start writes it as "a maximum output length of 128K tokens"; the Chat Completions reference puts the ceiling at 131072 and the default at 128k. One number, three ways of writing it, and the largest of them is the true ceiling.

• Input modalities — documented as text. The same Retrieve Model response returns input_modalities: ["text"] with a modality string of text->text. The changelog claims image input. These two things are on the same documentation site and they disagree; the next section deals with it.

• Output modalities — text, with the model's thinking returned separately in a reasoning_content field rather than woven into the answer. Reasoning tokens are counted separately in the usage block, which makes cost attribution on long chains tractable.

• Price per million tokens — $0.30 uncached input, $0.006 cached input, $1.20 output. The page labels all three "limited-time". The same table carries the renminbi rate card, ¥2 / ¥0.04 / ¥8, for accounts billed in China.

• Parameters — not published for this model. The 1.6T/48B shape is LongCat 2.0's.

• Independent score — none exists. Not a low score: no score.

Where the vendor contradicts itself about images

The changelog's headline claim is unambiguous: "Multimodal Understanding: New image understanding capability, able to parse image content, supporting cross-modal Q&A, content summarization, and complex visual reasoning." For a model whose only other documented differentiator is being cheap, that is the feature you would build on.

The API reference says the opposite. Retrieve Model, called for this exact model ID, returns input_modalities: ["text"] and output_modalities: ["text"]. The Chat Completions reference describes the message content field as "Message content as a plain-text string" — a string, not an array of content parts, with no image URL parameter and no attachment field anywhere in the documented request body. Nothing in the documented interface has a slot to put a picture in.

One of those two pages is stale, and there is no way to tell which from the outside. The changelog is a release announcement and announcements are written when a capability lands; the API reference is generated from what the endpoint actually accepts, and reference pages are the last thing anyone updates. That reasoning points one way, and it is reasoning rather than evidence, which is why this page will not assert image input as a working feature. If vision is the reason you are interested, spend ten minutes sending one image and reading the response. Do not plan a pipeline around the changelog wording, and do not rule the model out on the strength of the reference page either.

There is no benchmark. That is the state of it.

Meituan has published no benchmark for LongCat-2.5-Preview. There is no eval table in the documentation, no comparison chart in the changelog, no score on the pricing page, and no release blog post to carry one — Meituan's own model page for the previous generation has a technical write-up attached, and 2.5 has nothing equivalent that we could find on any Meituan surface. That is not an oversight in this article; it is the first fact a buyer needs.

Nor is there an independent score. Artificial Analysis, which does measure the LongCat line — it carries LongCat 2.0 with a release date of 29 June 2026, and LongCat Flash Lite — had no LongCat 2.5 entry in its model index when we checked on 13 October 2026. If you see a LongCat 2.5 benchmark number this month, the first question to ask is what it is actually attached to, because the most likely answer is that it is a LongCat 2.0 result relabelled by someone who assumed the two models share a profile. The parameter-count confusion described above is the same error in a different costume.

What the vendor does claim, all of it unreproduced, is the changelog's three-item list: multimodal understanding, "superior coding capability" spanning "code generation, code understanding, and automated programming tasks", and efficient operation with Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code. Treat those as the vendor's description of intent. Two of the three are claims every model card in this class makes, and the third — environment compatibility — is the kind of thing you can verify yourself in an afternoon by pointing your agent at the endpoint and seeing whether the tool calls come back in a shape your harness understands.

How to get a response out of it

The platform exposes two compatible surfaces over one key, issued from the API Keys page of the console and passed as Authorization: Bearer.

• OpenAI-compatible — POST https://api.longcat.chat/openai/v1/chat/completions. Point the Open​AI SDK's base_url at https://api.longcat.chat/openai and it works unchanged; the vendor's own example sends model: "LongCat-2.5-Preview" with max_tokens: 1000 and temperature: 0.7.

• Anthropic-compatible — POST https://api.longcat.chat/anthropic/v1/messages with the anthropic-version: 2023-06-01 header, and https://api.longcat.chat/anthropic as the Anthrop​ic SDK's base URL. Both request and response formats follow the Messages API.

• Accepted parameters — max_tokens up to 131072 (default 128k), temperature from 0 to 1, top_p, stream for SSE, tools and tool_choice, and a thinking object that takes {"type":"enabled"} or {"type":"disabled"}. Reasoning is a switch on the request, not a separate model ID.

• Rate limiting — an over-limit request returns HTTP 429 with retry_after: 60, and the docs recommend client-side exponential backoff. No requests-per-minute or tokens-per-minute ceiling is published anywhere, so you cannot size a concurrency budget from the documentation; you discover the limit by reaching it. For an agent loop that fires dozens of calls per task, that is worth knowing before the first 429 rather than after.

• Billing — prepaid balance, charged per request, on successful responses only. Failed requests — the documentation names 401, 403, 429 and 500 — are explicitly not billed. Balance does not expire, and promotional credit is consumed before paid balance.

Two details that will cost you a debugging session

The first is the reasoning field. The vendor's own response example returns an assistant message carrying both content and reasoning_content, with reasoning_tokens broken out inside usage.completion_tokens_details. That is not the Open​AI shape and it is not the Anthrop​ic shape: it is an interleaved field that arrives on every response. A client that deserialises into a fixed schema and drops unknown keys will quietly discard the model's thinking — and the failure is invisible, because the answer still comes back. You would conclude the model is weaker than it is, and you would be wrong for a reason that has nothing to do with the model. Check that your client keeps the field before you judge output quality.

The second is the modality question above. It is on this list because of how the two documents are distributed: the changelog is the page you read to understand what the model can do, and the API reference is the page you read to understand how to ask. Here they disagree about the same model, and the disagreement is exactly about the feature the release note leads with.

What would change this recommendation

This page is a snapshot, and it is a snapshot of a model with four open questions rather than a settled one. Five things would move it:

• A rename or a GA entry in the changelog — the moment -Preview stops being the model ID, the price stops being labelled limited-time, and a stability commitment becomes meaningful.

• A benchmark, from Meituan or from a measurement organisation — either would replace the current situation, which is that the model's competence is unknown rather than merely low.

• Image input demonstrated in the reference rather than asserted in the release note — the fix here is a one-line change to a generated page, and it would resolve the only internal contradiction in the documentation set.

• A parameter disclosure or a Hugging Face repository — either would let the 2.5 architecture be discussed on evidence instead of by inheritance from LongCat 2.0.

• A published rate limit — the smallest of the five, and the one that most affects whether an agent workload is plannable.

In the meantime the practical posture is straightforward, and it is not the same posture you would take with a scored, priced, stable model. An unscored preview at a promotional price is a model you evaluate, not one you point a production path at — and that makes the fallback the part worth setting up carefully. On OrcaRouter, the models we serve sit behind one key at provider list price with zero markup, so a vendor's price change reaches your bill the same day rather than at renewal, and automatic failover moves a degraded request to a healthy provider without your code knowing. LongCat-2.5-Preview is not one of our routes and we do not serve it, so trying it means going to Meituan's platform directly; the argument here is only that whatever you keep behind it should cost you a config line to reach, because the question this model poses — is it good enough for my work — is one you can only answer by running it, and the answer may well be no.

Which is the shape of the whole thing. LongCat 2.5 is a real model on a real platform with a documented interface, a published price and a 1M-token context window, and it is also a preview whose parameters, benchmark and vision support are all unestablished. Both halves are true, they are both worth writing down, and the second half is the one nobody republishes.