---
title: Meituan Releases Weights for LongCat-2.5 AI Agent Model
url: https://www.elseif.net/meituan-releases-weights-for-longcat-25-ai-agent-model
published: 2026-11-17T22:05:38+00:00
language: en
section: Models
source: https://the-decoder.de/longcat-2-5-meituan-gibt-gewichte-seines-16-billionen-parameter-modells-fuer-ki-agenten-frei/
organizations: Meituan, Huawei
publisher: elseif
---

# Meituan Releases Weights for LongCat-2.5 AI Agent Model

Meituan has released the weights for LongCat-2.5, a model designed for AI agents that can independently perform long, multi-step tasks such as programming. The model features 1.6 trillion parameters and was trained entirely on Chinese hardware.

LongCat-2.5 is a Mixture-of-Experts model that activates approximately 48 billion of its specialized subnetworks per processed token. It can process up to one million tokens at once via API and generate responses up to 128,000 tokens. The model understands text, images, and videos, providing responses in text form with a reasoning feature that can be toggled on or off.

Building on LongCat-2.0, the new version converts images into tokens that are processed alongside text, rather than using a separate image encoder. This approach, based on the LongCat-Next research model, is intended to improve capabilities in 3D tasks, design, and visual programming. To train the model, the team used distillation from specialized teacher models and reinforcement learning based on agent failures and real software tasks from GitHub pull requests. Training environments were restricted by removing cached files, Git history, and internet access to ensure tasks were solved legitimately.

In terms of performance, LongCat-2.5 scored 88.4 points on Terminal Bench v2.1, placing it slightly behind Claude Opus 5, which scored 89.1. On the more difficult Terminal Bench v4.0, it achieved 38.5 points, surpassing GPT-5.6 Sol at 37.3 but trailing Opus 5 at 53.9. On the DeepSWE benchmark, it scored 64.6 points, which is higher than Claude Opus 4.8 but lower than GPT-5.6 Sol and Deepseek-V4.1-Flash. The model achieved the highest score in the table for the BrowseComp web search task with 91.9 points, though this was conducted in a proprietary test environment.

Meituan provides a configuration that runs on 194 Huawei Ascend Atlas-A2 series accelerators, with 128 dedicated to generating responses, 64 to reading inputs, and two for image processing. The model uses 4-bit weights and intermediate values, which Meituan claims increases prompt reading throughput by 43.3 percent and response generation throughput by 42.77 percent compared to an 8-bit version without losing accuracy. No instructions for Nvidia GPUs are provided.

The weights are available under the MIT license. The API supports OpenAI and Anthropic formats and integrates with agent harnesses including Claude Code, Codex, and Hermes. Initial promotional pricing is approximately 0.27 Euro per million input tokens and 1.07 Euro per million output tokens.
