美团发布了 LongCat-2.0，这是一个大规模混合专家（MoE）语言模型。该模型拥有 1.6 万亿总参数量，每个 token 激活约 480 亿参数。它专为智能体编程场景设计：在智能体工作流中实现代码理解、生成与执行。
有两个事实尤为突出。首先，LongCat-2.0 支持原生 100 万 token 的上下文窗口。其次，其训练和推理服务完全运行在国产 AI ASIC 超级计算集群上。
什么是 LongCat-2.0？
LongCat-2.0 是美团推出的下一代万亿参数开源模型。它延续了 2025 年发布的 5600 亿参数模型 LongCat-Flash。其架构设计围绕一个核心目标：实现可靠、高效的智能体编程。
预训练过程消耗了超过 35 万亿个 token，累计使用数百万加速器小时。美团报告称，训练期间未出现回滚或不可恢复的损失尖峰。这一稳定性声明在工具链尚不成熟的非英伟达硬件上尤为重要。
架构：1.6T 模型如何保持低成本运行
该设计融合了四种降低规模化成本的思路，每种思路都值得单独理解。
- 零计算专家：并非每个 token 都需要大量计算。像标点符号这类简单 token 会被路由到零计算专家，原样返回。复杂 token 则会调用更多专家容量。一个 PID 控制器会调整专家偏置，使平均值保持在合理范围内。这产生了 330 亿至 560 亿参数的动态激活窗口，而非固定成本。MoE 主干采用捷径连接设计（ScMoE）以实现更高吞吐量。
- LongCat 稀疏注意力（LSA）：标准注意力的计算量随上下文长度呈二次方增长。LSA 仅选择最相关的 token，将复杂度降至接近线性。美团将其描述为 DeepSeek 稀疏注意力（DSA）的演进版本。它融合了三种正交的索引方法：流感知索引将碎片化的内存读取转换为连续块；跨层索引在相邻层之间复用注意力显著性；层级索引采用由粗到精的两阶段过滤。三者协同作用，使得 100 万 token 窗口得以维持，而不会遇到内存瓶颈。
- N-gram 嵌入：该设计新增了一个 1350 亿参数的 N-gram 嵌入模块。它在稀疏维度上与 MoE 专家模块正交排列。美团表示，该模块能够捕捉密集的局部 token 关系，同时在大批量解码过程中减少内存 I/O。
- 后训练（MOPD）：一条专用流水线（MOPD）融合了三组教师专家模块。这些模块分别涵盖智能体、推理和交互能力，最终整合为一个统一模型。
在服务部署方面，美团采用了六维并行方案和预填充-解码分离架构。同时还使用了“超级内核”和 L2 缓存权重预取技术来隐藏 I/O 延迟。
基准测试
美团将 LongCat-2.0 定位为智能体式编程模型。以下所有数据均来自美团内部测试。
| 基准测试 | LongCat-2.0 | 测试内容 |
|---|---|---|
| SWE-bench Pro | 59.5 | 真实世界软件工程任务 |
| Terminal-Bench 2.1 | 70.8 | Shell 环境中的执行与错误恢复 |
| SWE-bench Multilingual | 77.3 | 跨语言仓库任务 |
在 SWE-bench Pro 上，美团报告 LongCat-2.0 以微弱优势领先 GPT-5.5（58.6）。美团还声称其整体性能与 Google 的 Gemini 3.1 Pro 相当。所报告的领先优势主要集中在软件工程领域。在更广泛的通用智能体基准测试（如 FORTE 和 BrowseComp）中，覆盖范围表明它落后于领先的前沿系统。目前尚无独立排行榜的确认结果。
LongCat-2.0 对比 LongCat-Flash
从纸面数据来看，相比上一代产品有大幅提升。下表使用各模型已公布的规格参数。
| 属性 | LongCat-2.0 | LongCat-Flash |
|---|---|---|
| 总参数量 | 1.6T | 560B |
| 每个 token 激活参数量 | 约 48B（33B–56B） | 约 27B（18.6B–31.3B） |
| 上下文窗口 | 100 万 token（原生） | 128K token |
| 长上下文注意力机制 | LongCat 稀疏注意力 | 多头潜在注意力 |
| 报告所用硬件 | 国产 AI ASIC 超级计算集群（训练 + 服务部署） | H800 GPU（报告用于推理） |
| 最大输出 | 128K token | 未指定 |
| 许可证 | MIT | MIT |
| 发布时间 | 2026 年 6 月 30 日 | 2025 年 9 月 |
| 权重 | 即将发布 | 已开源 |
使用场景与示例
LongCat-2.0 针对智能体式的软件工作进行了调优，而非用于日常闲聊。以下几个具体模式能够发挥其优势。
- 全仓库推理：将整个中等规模的代码库喂入百万 token 上下文窗口。让模型一次性跨多个文件追踪 bug。这避免了短上下文窗口所迫使用的摘要式取巧方法。
- 多步骤终端任务：在具备 shell 访问权限的智能体循环中运行模型。它可以执行命令、读取错误信息并重试，直到任务通过。Terminal-Bench 2.1 基准测试正是针对这一工作流。
- 仓库级编辑：要求对跨多个模块和测试的代码进行重构。模型在提出协调一致的修改方案之前，会基于完整上下文进行推理。
- 跨语言迁移：利用 SWE-bench Multilingual 基准测试的优势处理多语言仓库。模型可以在不同语言之间移植逻辑，同时保持行为一致。
这些模式在标准智能体框架内运行。因此，开发团队无需构建新工具即可采用该模型。
如何获取
LongCat-2.0 可通过 LongCat API 平台访问。它同时提供兼容 OpenAI 和兼容 Anthropic 的端点。该模型也上线了 OpenRouter，并可在 Claude Code、OpenClaw、OpenCode 和 Codex 等框架中使用。目前尚不支持本地自托管，因为权重尚未发布。
兼容 OpenAI 的端点使用模型 ID LongCat-2.0。最大输出长度为 131072 个 token（128K）。以下代码片段调用了文档中描述的聊天补全端点。
# pip install openai
from openai import OpenAI
client = OpenAI(
api_key="YOUR_LONGCAT_API_KEY",
base_url="https://api.longcat.chat/openai/v1",
)
resp = client.chat.completions.create(
model="LongCat-2.0",
messages=[
{"role": "system", "content": "You are a coding agent."},
{"role": "user", "content": "Refactor utils.py to remove duplicate I/O logic."},
],
max_tokens=4096, # LongCat-2.0 supports up to 131072 (128K)
)
print(resp.choices[0].message.content)定价为每百万输入 token 0.75 美元，每百万输出 token 2.95 美元。发布促销价为 0.30 美元和 1.20 美元，缓存上下文读取免费。这些数据来自第三方报道，可能发生变化。
交互式说明
核心要点
- 以 MIT 许可证发布
- LongCat-2.0 是一个 1.6 万亿参数的 MoE 模型，每个 token 激活约 480 亿参数（动态范围 330 亿至 560 亿）。
- 原生百万 token 上下文来自 LongCat 稀疏注意力机制，将长上下文成本从二次方降至线性。
- 训练和推理在 5 万张国产 AI ASIC 集群上运行，未使用英伟达硬件。
- 厂商报告的得分：SWE-bench Pro 59.5 分，Terminal-Bench 2.1 70.8 分，SWE-bench Multilingual 77.3 分。
Meituan has released LongCat-2.0, a large-scale Mixture-of-Experts (MoE) language model. It carries 1.6 trillion total parameters and activates about 48 billion per token. The model targets agentic coding: code understanding, generation, and execution inside agent workflows.
Two facts stand out. First, LongCat-2.0 supports a native 1-million-token context window. Second, both training and serving ran entirely on domestic AI ASIC superpods.
What is LongCat-2.0?
LongCat-2.0 is Meituan’s next-generation trillion-parameter open model. It follows LongCat-Flash, a 560B model released in 2025. The architecture was designed around one goal: reliable, efficient agentic coding.
Pretraining spanned more than 35 trillion tokens over millions of accelerator-hours. Meituan reports no rollbacks or irrecoverable loss spikes during the run. That stability claim matters on non-Nvidia hardware, where tooling is less mature.
Architecture: How a 1.6T Model Stays Cheap to Run
The design combines four ideas that reduce the cost of scale. Each one is worth understanding on its own.
- Zero-computation experts: Not every token needs heavy compute. Simple tokens like punctuation route to a zero-computation expert and return unchanged. Complex tokens engage more expert capacity. A PID controller adjusts expert bias to hold the average in range. This produces the 33B–56B dynamic activation window instead of a fixed cost. The MoE backbone uses a shortcut-connected design (ScMoE) for higher throughput.
- LongCat Sparse Attention (LSA): Standard attention scales quadratically with context length. LSA selects only the most relevant tokens, dropping the scaling closer to linear. Meituan describes it as an evolution of DeepSeek Sparse Attention (DSA). It layers three orthogonal indexing methods. Streaming-aware Indexing turns fragmented memory reads into contiguous blocks. Cross-Layer Indexing reuses attention saliency across adjacent layers. Hierarchical Indexing applies coarse-to-fine two-stage filtering. Together they sustain the 1M-token window without a memory wall.
- N-gram Embedding: The design adds a 135-billion-parameter N-gram embedding module. It sits orthogonal to the MoE experts in sparse dimensions. Meituan says it captures dense local token relationships. It also reduces memory I/O during large-batch decoding.
- Post-training (MOPD): A dedicated pipeline (MOPD) fuses three teacher expert groups. These cover Agent, Reasoning, and Interaction capabilities into one unified model.
For serving, Meituan uses a 6D parallelism scheme and a prefill-decode disaggregated architecture. It also employs ‘super kernels’ and L2-cache weight prefetching to hide I/O latency.
Benchmarks
Meituan positions LongCat-2.0 as an agentic coding model. Every figure below comes from Meituan’s own testing.
| Benchmark | LongCat-2.0 | What it measures |
|---|---|---|
| SWE-bench Pro | 59.5 | Real-world software engineering tasks |
| Terminal-Bench 2.1 | 70.8 | Execution and error recovery in shells |
| SWE-bench Multilingual | 77.3 | Cross-language repository tasks |
On SWE-bench Pro, Meituan reports LongCat-2.0 edging GPT-5.5 (58.6). Meituan also claims overall performance comparable to Google’s Gemini 3.1 Pro. The reported edge is concentrated in software engineering. On broader general-agent benchmarks such as FORTE and BrowseComp, coverage indicates it trails leading frontier systems. Independent leaderboard confirmation is not yet available.
LongCat-2.0 vs LongCat-Flash
The jump from the previous generation is large on paper. This table uses each model’s published specifications.
| Attribute | LongCat-2.0 | LongCat-Flash |
|---|---|---|
| Total parameters | 1.6T | 560B |
| Active per token | ~48B (33B–56B) | ~27B (18.6B–31.3B) |
| Context window | 1M tokens (native) | 128K tokens |
| Long-context attention | LongCat Sparse Attention | Multi-head Latent Attention |
| Reported hardware | Domestic AI ASIC superpods (training + serving) | H800 GPUs (inference reported) |
| Max output | 128K tokens | Not specified |
| License | MIT | MIT |
| Released | June 30, 2026 | September 2025 |
| Weights | Coming soon | Open |
Use Cases With Examples
LongCat-2.0 is tuned for agent-style software work, not casual chat. A few concrete patterns fit its strengths.
- Whole-repository reasoning: Feed an entire mid-sized codebase into the 1M-token window. Ask the model to trace a bug across many files at once. This avoids the summarization hacks that shorter windows force.
- Multi-step terminal tasks: Run the model inside an agent loop with shell access. It can execute commands, read errors, and retry until a task passes. The Terminal-Bench 2.1 focus targets exactly this workflow.
- Repository-level edits: Ask for a refactor that spans several modules and tests. The model reasons over the full context before proposing coordinated changes.
- Cross-language migration: Use the SWE-bench Multilingual strength for polyglot repositories. The model can port logic between languages while preserving behavior.
These patterns run inside standard agent harnesses. Dev teams can therefore adopt the model without building new tooling.
How to Access It
LongCat-2.0 is reachable through the LongCat API Platform. It exposes both OpenAI-compatible and Anthropic-compatible endpoints. The model is also on OpenRouter and in harnesses like Claude Code, OpenClaw, OpenCode, and Codex. Local self-hosting is not yet possible, since weights remain pending.
The OpenAI-compatible endpoint uses the model ID LongCat-2.0. Maximum output length is 131072 tokens (128K). The snippet below calls the documented chat-completions endpoint.
# pip install openai
from openai import OpenAI
client = OpenAI(
api_key="YOUR_LONGCAT_API_KEY",
base_url="https://api.longcat.chat/openai/v1",
)
resp = client.chat.completions.create(
model="LongCat-2.0",
messages=[
{"role": "system", "content": "You are a coding agent."},
{"role": "user", "content": "Refactor utils.py to remove duplicate I/O logic."},
],
max_tokens=4096, # LongCat-2.0 supports up to 131072 (128K)
)
print(resp.choices[0].message.content)Pricing is reported at $0.75 per million input tokens and $2.95 per million output. A launch promotion lists $0.30 and $1.20, with cached context reads free. These figures come from third-party coverage and may change.
Interactive Explainer
Key Takeaways
- Released under MIT
- LongCat-2.0 is a 1.6T-parameter MoE model activating ~48B parameters per token (dynamic range 33B–56B).
- Native 1M-token context comes from LongCat Sparse Attention, cutting long-context cost from quadratic to linear.
- Training and inference ran on a 50,000-card domestic AI ASIC cluster, with no Nvidia hardware.
- Vendor-reported scores: 59.5 SWE-bench Pro, 70.8 Terminal-Bench 2.1, 77.3 SWE-bench Multilingual.