Skip to content
Theme
OnDevice Tools

AIAI Models & Research

Meituan releases LongCat-2.5 agent model trained without Nvidia GPUs

Chinese delivery giant Meituan has open-sourced LongCat-2.5, a 1.6-trillion-parameter model for AI agents trained entirely on domestic hardware. The model uses a mixture-of-experts architecture and achieves competitive scores against top Western models on coding benchmarks, though it trails slightly

Diagram of LongCat-2.5 architecture trained on Huawei Ascend chips

Key takeaways

  • Meituan has released LongCat-2.5, a 1.6-trillion-parameter AI agent model trained entirely on Chinese hardware with no Nvidia GPUs.
  • The model uses a mixture-of-experts architecture with 48 billion active parameters and achieves benchmark scores close to top Western models like Claude Opus 5.
  • A sample deployment runs on 194 Huawei Ascend chips, and the model uses 4-bit quantization to improve throughput on Chinese hardware.
  • The weights are available under the MIT license for commercial use, with a promotional API pricing of about 27 euro cents per million input tokens.
In this article
  1. What is LongCat-2.5?
  2. Training on Chinese hardware
  3. Benchmark performance
  4. Availability and pricing
  5. Broader context

What is LongCat-2.5?

Illustration: AI, code, crypto, NVIDIA logo and the labels AI Agents, Claude, Huawei Ascend

Meituan, the Chinese food delivery and services platform, has released the weights for LongCat-2.5 this week, a 1.6-trillion-parameter model designed for AI agents. According to The Decoder, the model was trained entirely on Chinese-made hardware, with no Nvidia GPUs involved. LongCat-2.5 is built to handle long, multi-step autonomous tasks, with a particular emphasis on coding and software development.

The model uses a mixture-of-experts architecture, where many specialized subnetworks exist but only a portion are activated for each token processed. The model has 1.6 trillion parameters in total, but only about 48 billion are active when processing any given token. Through its API, the model can process up to 1 million tokens at a time and produce outputs of up to 128,000 tokens. It understands text, images, and video, and produces text output. Users can also switch reasoning on or off depending on their needs.

LongCat-2.5 is based on the company's earlier LongCat-2.0 model and adds image understanding. Instead of adding a standalone image encoder later on, as many multimodal models do, LongCat-2.5 transforms images into tokens that the model processes along with text. This approach, drawn from Meituan's research model LongCat-Next, aims to improve performance on visual coding, design, and 3D tasks.

Training on Chinese hardware

One of the most notable aspects of this release is the hardware used for training. Meituan states that it trained the earlier LongCat-2.0 model in late June using a cluster of over 50,000 domestic AI chips. For LongCat-2.5, the company has published instructions for running the model on 194 Huawei Ascend chips from the Atlas A2 line.

The deployment splits work across these chips in a specific way. A pair of them handles image processing tasks. Another group manages prompt ingestion, and a third group handles response generation. The model uses 4-bit quantization, a format that Chinese chips can handle without special conversion. According to Meituan's testing, this approach boosted prompt processing throughput by 43.3 percent and lifted response generation throughput by 42.77 percent relative to an 8-bit setup, while keeping accuracy equivalent to the unquantized version.

Benchmark performance

LongCat-2.5 achieves competitive scores on several agent benchmarks. On Terminal Bench v2.1, which measures how well models solve command-line tasks, it scores 88.4 points. This puts it slightly behind Claude Opus 5 with 89.1, GPT-5.6 Sol at 88.8, and Deepseek-V4.1-Flash with 90.6.

On DeepSWE, which evaluates how well models repair bugs in actual codebases, LongCat-2.5 scores 64.6 points, a substantial improvement from its predecessor's 9.7. This puts it behind Deepseek-V4.1-Flash, which scored 74.2, and GPT-5.6 Sol at 72.7, but ahead of several other models. Claude Opus 5.5 and GPT-6.1 Sol, which are more recent, are not included in the comparison table.

On the more difficult Terminal Bench v4.0, LongCat-2.5 achieves 38.5 points, just above GPT-5.6 Sol's 37.3 but far from Claude Opus 5's 53.9. On the web search benchmark BrowseComp, LongCat-2.5 posts 91.9 points, the highest score in Meituan's comparison table. However, Meituan conducted this test using its own infrastructure, so the result should be treated with some caution.

Availability and pricing

The LongCat-2.5 weights are released under the MIT license, which permits commercial use, modification, and redistribution. Users can also try the model through a chat interface at longcat.ai. The API is compatible with the OpenAI and Anthropic formats, and it integrates with agent frameworks such as Claude Code, Codex, and Hermes.

When the service launches, Meituan is offering a limited-time promotional rate of about 27 euro cents per million input tokens and 1.07 euro per million output tokens.

Broader context

The release comes at a time when Chinese open-weight models are rapidly closing the gap with their American counterparts. Chinese open-weight models, including Kimi K3.1 and GLM-5.3, have nearly matched top American models on standard benchmarks. Meituan's focus on Chinese hardware from the start makes that infrastructure more attractive to potential adopters outside China as well.

Beyond agent models, Meituan is also developing image generation tools. The company previously released LongCat-Image, a text-to-image model available under an open-source license. It uses only 6 billion parameters and emphasises efficiency over raw scale.

Sources

  1. Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPUThe Decoder

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.