美团发布 1.6 万亿参数智能体模型 LongCat-2.5 权重,训练全程未用 Nvidia GPU
Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU
- 1美团开源 LongCat-2.5 权重,1.6 万亿参数智能体模型,官方称训练全程只用国产芯片。
- 2官方基准称接近 Claude Opus 5,但 DeepSWE 修 Bug 测试仅 64.6 分,排名靠后。
- 3官方示例部署需 194 张华为昇腾 Atlas A2 加速卡,文档未提供 Nvidia GPU 方案。
- 4API 支持百万 token 上下文,限时价约 €0.27 每百万输入 token,可接入 Claude Code、Codex。
美团用纯国产芯片训练并开源了万亿参数智能体模型,无 Nvidia 路线从自用走向可被外部复用。
- 影响:企业多了一个不依赖 Nvidia 的开放权重智能体选项,MIT 协议允许商用和再分发;API 限时促销价约 €0.27 每百万输入 token,智能体调用成本处在低位;但 DeepSWE 等编码基准仍落后一线模型,能力边界要按任务核对。
- 参考:可考虑用其 API 在 Claude Code 或 Codex 里跑一组真实编码与检索任务做对照测试;若评估国产算力路线,值得参照 194 张昇腾卡的官方自托管配置先估算集群成本。
- 依据:基准分数、芯片配置和价格均为美团官方口径,暂无独立验证;BrowseComp 91.9 分系美团自测环境,且基准表未含更新的 Claude Opus 5.5 与 GPT-6.1 Sol。
Meituan releases the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents. It was trained entirely on Chinese hardware and performs at the level US models like Claude Opus 5 reached a few months ago on coding benchmarks.
The LongCat team at Chinese food delivery company Meituan has released the weights for LongCat-2.5, one week after the model's official unveiling. LongCat-2.5 is built for agents that work through long, multi-step tasks autonomously, particularly coding. According to Meituan, it comes close to top Western models on several benchmarks.
LongCat-2.5 is a mixture-of-experts model that activates only a subset of its specialized subnetworks for each token it processes. Out of 1.6 trillion total parameters, roughly 48 billion are active per token. Through the API, the model handles up to one million tokens at once and generates responses up to 128,000 tokens long. It understands text, images, and video, responds in text, and lets users toggle reasoning on and off.
LongCat-2.5 converts images into tokens
LongCat-2.5 builds on its predecessor LongCat-2.0 and adds image understanding. Many multimodal models bolt on a separate image encoder after the fact. LongCat-2.5 instead converts images into tokens that the model processes alongside text, as is standard practice for today's LLMs. The approach draws on the research model LongCat-Next and aims to improve visual coding, design, and 3D tasks.

For reinforcement learning, the team used real software tasks from GitHub pull requests along with recreated agent failures. To make sure the model actually solved the tasks during training, the developers blocked internet access in the training environments, removed Git history and cached files, and had a language model check its solutions for tricks. The team also distilled the capabilities of several specialized teacher models for agent work, reasoning, and dialog into LongCat-2.5.

A sample deployment runs on 194 Huawei chips
Meituan quantizes the experts' weights and activations to 4-bit precision, which Chinese chips can process natively. In Meituan's load tests, throughput rises 43.3 percent for prompt processing and 42.77 percent for response generation compared with an 8-bit version. Meituan says accuracy stays on par with the uncompressed model.
This version is available in a separate repository. The self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series. Two handle image processing, 64 handle prompt ingestion, and 128 handle response generation. There's no guidance for Nvidia GPUs.
LongCat-2.5 trails Opus 5 slightly on terminal tasks
The LongCat team reports 88.4 points on Terminal Bench v2.1, a benchmark that measures how well agents solve command-line tasks. Meituan lists Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8, and Deepseek-V4.1-Flash at the top of the table with 90.6.
DeepSWE tests how well models fix bugs in real code repositories. Here LongCat-2.5 jumps from 9.7 to 64.6 points, placing it behind every other model in the table except Claude Opus 4.8, with Deepseek-V4.1-Flash leading at 74.2 and GPT-5.6 Sol at 72.7. The more recent Claude Opus 5.5 and GPT-6.1 Sol are not included in the table.

Terminal Bench v4.0, the harder variant, puts LongCat-2.5 at 38.5 points, slightly above GPT-5.6 Sol's 37.3 but well below Opus 5's 53.9. The model posts 91.9 on the web search benchmark BrowseComp, the highest score in the table, though Meituan ran this test in its own environment.
The weights are available under the MIT license, which allows commercial use, modification, and redistribution. Users can also try LongCat-2.5 in a chat interface at https://longcat.ai/. Its API supports both the OpenAI and Anthropic formats and plugs into agent harnesses like Claude Code, Codex, and Hermes. At launch, Meituan is charging a limited-time promotional price equivalent to roughly €0.27 per million input tokens and €1.07 per million output tokens.
Meituan says it trained the predecessor LongCat-2.0 in late June entirely on a cluster of more than 50,000 Chinese AI chips. Before that, the model ran anonymously as "Owl Alpha" on OpenRouter. Chinese open-weight models like Kimi K3.1 and GLM-5.3 have largely caught up with top US models on common benchmarks.
Open-weight models optimized for Chinese chips from the start make that hardware more attractive outside China too. Beyond multimodal agent models, Meituan is also working on image generators. The company previously released LongCat-Image, an open-source text-to-image model that gets by with just 6 billion parameters and prioritizes efficiency over sheer model size.
来源:The Decoder:AI News · the-decoder.com