Meituan Releases Open LongCat-2.5 Agent Model Trained on Huawei Hardware
Meituan releases LongCat-2.5, a 1.6-trillion-parameter MoE agent model trained entirely on Chinese hardware that rivals top Western systems on coding and terminal benchmarks.

Chinese technology company Meituan has released the open-source weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model optimized for autonomous AI agents and coding tasks. Arriving one week after its official unveiling, the model was trained completely without Nvidia GPUs, relying instead on domestic Chinese hardware. Designed to handle long-context, multi-step engineering workflows, LongCat-2.5 demonstrates that high-parameter agent models can achieve benchmark results close to proprietary Western systems while operating on alternative hardware architectures.
Model Architecture and Hardware Infrastructure
LongCat-2.5 uses a mixture-of-experts (MoE) structure that activates roughly 48 billion of its 1.6 trillion total parameters per token. Building upon LongCat-2.0 and the research prototype LongCat-Next, the model supports text, image, and video inputs, generating text outputs while offering toggleable step-by-step reasoning. Instead of utilizing a separate vision encoder, LongCat-2.5 converts native-resolution images directly into discrete tokens processed alongside text, aiming to improve visual coding, 3D operations, and design tasks. Via API, the model handles context windows up to one million tokens and can generate responses up to 128,000 tokens long.
For execution, Meituan provides 4-bit native quantization for the model's expert weights and activations, a format tailored for Chinese hardware. In company load tests, 4-bit quantization increased prompt throughput by 43.3 percent and response generation by 42.77 percent compared to an 8-bit deployment, while maintaining accuracy. Meituan's hosting deployment guide details a cluster of 194 Huawei Ascend Atlas A2 series accelerators—using two chips for image processing, 64 for prompt ingestion, and 128 for output generation—without providing setup guidance for Nvidia GPUs.
Agent Training Methods and Benchmark Performance
To prepare LongCat-2.5 for autonomous task execution, Meituan trained the model using real software tasks extracted from GitHub pull requests and recreated agent failure scenarios. Training sandboxes strictly blocked internet access, erased Git histories, and cleared cached files, while an auxiliary language model verified solutions to ensure no shortcuts were taken. Meituan also distilled capabilities into LongCat-2.5 from several specialized teacher models focused on reasoning, dialogue, and agent execution.
In standard evaluations, LongCat-2.5 earned 88.4 points on Terminal Bench v2.1, a command-line task benchmark. This places it close behind Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8, while Deepseek-V4.1-Flash leads at 90.6. On the harder Terminal Bench v4.0, LongCat-2.5 scored 38.5 points, placing above GPT-5.6 Sol (37.3) but below Claude Opus 5 (53.9). On DeepSWE, which tests automated bug fixing in software repositories, LongCat-2.5 scored 64.6 points—a substantial rise from LongCat-2.0's 9.7 points—ranking behind GPT-5.6 Sol (72.7) and Deepseek-V4.1-Flash (74.2). In Meituan's internal testing environment, the model achieved a top score of 91.9 on the BrowseComp web search benchmark.
What it means for developers
Released under the permissive MIT license, LongCat-2.5 allows developers to freely modify, commercially deploy, and self-host a large-scale agent model. Because its API adheres to both OpenAI and Anthropic request formats, engineering teams can integrate LongCat-2.5 directly into agent frameworks such as Claude Code, Codex, and Hermes.
Furthermore, the model demonstrates the operational viability of non-Nvidia hardware clusters alongside open-weight alternatives like Kimi K3.1 and GLM-5.3. Developers interested in evaluating different model architectures can try top AI models cheaply through one API at https://apixoai.online, streamlining integration and comparative testing. For direct usage, Meituan offers web access at https://longcat.ai/ alongside promotional API pricing of approximately €0.27 per million input tokens and €1.07 per million output tokens.
Source: Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key

