TAU-HOME.COM
LOADING

Meituan Releases 1.6-Trillion-Parameter Agent Model Trained Without Nvidia GPUs: LongCat-2.5 Open-Sourced Under MIT License

Chinese delivery giant Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter AI agent model trained entirely on Chinese hardware without Nv

tau · November 21, 2026

#Meituan #LongCat-2.5 #AIAgents #ChineseAI #HuaweiAscend #OpenSourceModels

Meituan Releases 1.6-Trillion-Parameter Agent Model Trained Without Nvidia GPUs: LongCat-2.5 Open-Sourced Under MIT License

The LongCat team at Chinese food delivery and local services company Meituan has officially released the weights for LongCat-2.5, a massive 1.6-trillion-parameter AI agent model. Notably, the model was trained entirely on domestic Chinese hardware without touching a single Nvidia graphics processing unit (GPU). The release is governed by the permissive MIT license, permitting commercial use, modification, and redistribution.

Meituan LongCat-2.5 AI model overview and Huawei Ascend cluster deployment architecture

Image source: https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu

LongCat-2.5 is built specifically for autonomous AI agents that perform long, multi-step tasks, with a strong emphasis on coding and software engineering. According to Meituan's internal and benchmark evaluations, the model approaches the performance levels achieved by top Western foundation models on several key tasks, illustrating how hardware-constrained environments are rapidly advancing independent AI infrastructures.

Training Entirely off Nvidia Hardware and Deploying on Huawei Ascend Clusters

The most striking aspect of the LongCat-2.5 release is its complete departure from Western hardware ecosystems. Following the predecessor LongCat-2.0—which was trained in late June on a cluster of over 50,000 Chinese AI chips after previously operating anonymously under the alias "Owl Alpha" on OpenRouter—this iteration further solidifies the viability of domestic Chinese hardware stacks.

The self-hosting and deployment documentation outlines a sample production cluster comprising 194 Huawei Ascend accelerators from the Atlas A2 series:

  • Cluster Workload Distribution: Out of the 194 chips, 2 are dedicated to image processing, 64 handle prompt ingestion and tokenization, and 128 manage response generation.
  • Quantization and Throughput: The model utilizes a mixture-of-experts (MoE) architecture where roughly 48 billion out of the 1.6 trillion total parameters are active per token. Meituan quantizes the experts' weights and activations down to 4-bit precision, enabling native execution on Chinese accelerators. In load tests, this quantization boosts throughput by 43.3% for prompt processing and 42.77% for response generation compared to an 8-bit baseline, while keeping accuracy on par with the uncompressed model.

Multimodal Tokenization and Rigorous Sandbox Reinforcement Learning

LongCat-2.5 expands upon its predecessor by integrating native image understanding alongside text. Rather than bolting on a separate image encoder post-hoc, LongCat-2.5 breaks down images at their native resolution into discrete tokens that are processed directly alongside text tokens, facilitating improved visual coding, user interface design, and 3D task handling.

To train the model for reliable agentic behavior via reinforcement learning, Meituan established a heavily restricted sandbox environment:

  • Shortcut Prevention: The training pipeline completely blocked internet access, stripped away Git history, and removed cached files from real GitHub pull request tasks. A separate language model verified that the agent actually solved tasks rather than exploiting shortcuts or web searches.
  • Model Distillation: The training process also distilled capabilities from multiple specialized teacher models focused on agent execution, multi-turn reasoning, and dialogue.

Benchmark Performance and API Accessibility

In published evaluations, LongCat-2.5 scores 88.4 on Terminal Bench v2.1, a benchmark measuring command-line task execution, putting it in a comparable tier with models like Claude Opus 5 (89.1) and GPT-5.6 Sol (88.8). However, on more rigorous coding bug-fix evaluations like DeepSWE, it trails behind leading Western flash models.

For developer integration, LongCat-2.5's API supports both OpenAI and Anthropic request formats, allowing it to plug directly into standard agent harnesses such as Claude Code, Codex, and Hermes. Alongside the open-weight release under the MIT license, Meituan offers a live chat testing interface at longcat.ai and has introduced an introductory promotional API pricing tier roughly equivalent to €0.27 per million input tokens and €1.07 per million output tokens.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285