Is LongCat-2.5 state of the art?Pending benchmarks.Announced Nov 1, 2026; no independent results yet.

The details so far

Lab

Meituan

Announced

Nov 1, 2026

API

Live · LongCat-2.5 · LongCat API Platform, model LongCat-2.5, at api.longcat.chat with OpenAI Chat Completions, OpenAI Responses and Anthropic Messages compatible endpoints; listed as available from 2026-11-02 in the API changelog

Open weights

Announced, not yet published · Not yet published: the blog lists GitHub and Hugging Face as coming soon, and the meituan-longcat Hugging Face org had no LongCat-2.5 repo as of Nov 2, 2026

Parameters

1.6T total, about 48B active per token (mixture of experts)

Type

Native multimodal model for long-horizon agentic tasks: text, image and video in, text out; thinking can be switched on or off

Architecture

Mixture of experts, 1.6T total and about 48B active per token; visual input is represented as multi-level discrete tokens

Context

1,048,576 tokens (1M), up to 262,144 output tokens (256K), per the LongCat API docs

Training

Builds on LongCat-2.0 with long-context training, large-scale long-horizon agentic reinforcement learning and multi-teacher distillation; trained entirely on domestic Chinese AI accelerators, per Meituan

Inference

W4A4 quantized serving on domestic AI accelerators, per the LongCat blog

Price

$0.30 per 1M input tokens, $0.006 per 1M cached input tokens and $1.20 per 1M output tokens (¥2, ¥0.04 and ¥8), labelled a limited-time discounted price by LongCat; prices are subject to change

Availability

LongCat chat at longcat.ai and the LongCat API

Earlier preview

LongCat-2.5-Preview launched on the LongCat API on Sep 25, 2026 (image understanding and coding), per the API changelog; a separate earlier release, not tracked here

Predecessor

LongCat-2.0, released on the LongCat API on Jun 30, 2026, per the API changelog

Tags marked unsure are not stated on the lab's own pages.

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

• Publicly announced with a source
• Usable by the public (API or weights)
• Listed on an independent board we track
• Enough results for a rank in a category
• Holds across two snapshots (crowns only)

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: 88.4 on Terminal Bench v2.1, 38.5 on Terminal Bench v4.0, 64.6 on DeepSWE v1.1, 60.5 on NL2Repo and 82.1 on PaperBench in LongCat's own evaluation.

Lab claim: 91.9 on BrowseComp, 63.1 on HLE with tools (text-only subset), 83.1 on MCP Atlas and 51.8 on AutomationBench v1.0.6.

Lab claim: 78.6 on Vision2Web and 85.2 on VideoMME.

Lab claim: Meituan says its scores on several mainstream benchmarks fall between Claude-Opus-4.8 and Claude-Opus-5.

Lab claim: W4A4 serving raises prefill throughput by 43.3% and decode throughput by 42.77% and cuts TTFT by 15.91% and TPOT by 19.72% against a W8A8 INT8 baseline.

Sources