Meituan Implementation of LongCat-2.0 Utilizing Domestic Compute Infrastructure
美團利用國產計算基礎設施實現 LongCat-2.0 部署
Introduction
Meituan has released LongCat-2.0, a large language model developed exclusively using Chinese hardware for both training and inference.
美團發佈了 LongCat-2.0,這是一個在訓練與推理階段均完全使用中國硬體開發的大語言模型。
Main Body
The deployment of LongCat-2.0 represents a significant shift in the technical trajectory of Chinese artificial intelligence. Characterized by 1.6 trillion parameters and a context window of one million tokens, the model utilizes a Mixture-of-Experts (MoE) architecture with approximately 48 billion active parameters. Unlike previous domestic iterations, such as DeepSeek V4-pro, which utilized local silicon solely for inference, LongCat-2.0 completed the computationally intensive pre-training phase on a cluster of 50,000 domestic accelerators. This was achieved through the utilization of large-scale ASIC superpods and Huawei's Collective Communication Library (HCCL) to facilitate inter-chip coordination.
LongCat-2.0 的部署代表了中國人工智慧技術軌跡的一個重大轉向。該模型具有 1.6 兆個參數以及一百萬個 token 的上下文視窗,採用專家混合 (MoE) 架構,約有 480 億個激活參數。與先前僅將國產晶片用於推理的國內版本(如 DeepSeek V4-pro)不同,LongCat-2.0 在一個由 5 萬個國產加速器組成的集群上,完成了運算密集型的預訓練階段。這是透過使用大規模 ASIC superpods 與華為的 Collective Communication Library (HCCL) 來促進晶片間的協調而實現的。
From a performance perspective, the model demonstrates parity with Google's Gemini 3.1 Pro, specifically outperforming it on Terminal-Bench 2.1 and SWE-Bench Pro. However, a performance gap persists when compared to frontier systems such as OpenAI's GPT-5.5 and Anthropic's Opus 4.8 in complex reasoning tasks. The development process revealed systemic vulnerabilities; Meituan noted that the domestic software ecosystem remains underdeveloped relative to NVIDIA's framework, and memory constraints per device were identified as a primary bottleneck during the training phase.
從性能角度來看,該模型表現與 Google 的 Gemini 3.1 Pro 持平,特別是在 Terminal-Bench 2.1 和 SWE-Bench Pro 上超越了後者。然而,在複雜推理任務中,與 OpenAI 的 GPT-5.5 和 Anthropic 的 Opus 4.8 等前沿系統相比,仍存在性能差距。開發過程揭露了系統性脆弱之處;美團指出,國產軟體生態相對於 NVIDIA 的框架仍不夠成熟,且每台設備的記憶體限制被確定為訓練階段的主要瓶頸。
Strategically, this development occurs within the context of United States export restrictions on high-end semiconductors. The successful scaling of a trillion-parameter model on local silicon suggests a diminishing reliance on prohibited hardware. This technical milestone is further supported by Meituan's strategic investments in domestic semiconductor firms, including MetaX and Moore Threads, indicating a broader institutional commitment to compute self-sufficiency.
在戰略上,此次開發發生在美國限制高端半導體出口的背景下。在國產晶片上成功擴展兆級參數模型,表明對禁運硬體的依賴程度正在降低。這一技術里程碑進一步得到了美團對國產半導體公司(包括 MetaX 和摩爾線程)戰略投資的支持,顯示出機構對實現計算自給自足的更廣泛承諾。
Conclusion
LongCat-2.0 establishes the technical viability of frontier-scale AI training on domestic Chinese hardware, despite remaining software and memory limitations.
儘管仍存在軟體與記憶體限制,但 LongCat-2.0 證明了在國產硬體上進行前沿規模 AI 訓練的技術可行性。
Vocabulary Learning
The Architecture of 'Nominalization' as a C2 Power-Tool
To bridge the gap from B2 to C2, a student must move away from action-oriented prose (subject verb object) and embrace concept-oriented prose. The provided text is a masterclass in Nominalization: the process of turning verbs or adjectives into nouns to create a dense, objective, and academic register.
🔍 The Linguistic Pivot
Look at the strategic transformation of ideas in the text. A B2 learner describes a process; a C2 writer describes a phenomenon.
- B2 Approach (Clausal): Meituan developed LongCat-2.0 and this shifted how Chinese AI develops technically.
- C2 Approach (Nominalized): "The deployment of LongCat-2.0 represents a significant shift in the technical trajectory of Chinese artificial intelligence."
By converting the action ("shifted") into a noun ("shift") and the process ("how it develops") into a complex noun phrase ("technical trajectory"), the writer achieves two things:
- Increased Information Density: More conceptual weight is packed into a single sentence.
- Objective Distance: The focus shifts from the actor (Meituan) to the implication (the shift in trajectory).
🛠️ Advanced Synthesis: The 'Noun + Prepositional Chain'
C2 mastery is signaled by the ability to chain nouns and modifiers to create highly specific meanings without relying on repetitive verbs.
*"...a diminishing reliance on prohibited hardware."
Instead of saying "they rely less on hardware that is prohibited," the author uses a nominal head (reliance) modified by a participial adjective (diminishing) and a prepositional qualifier (on prohibited hardware).
⚡ Scholarly Application: From Process to Concept
To implement this, replace your primary verbs with their noun counterparts and support them with a 'state-of-being' verb (e.g., represents, indicates, constitutes, signifies).
| B2 Verb-Centric | C2 Nominalized Transformation |
|---|---|
| Because the software is underdeveloped... | Due to the underdevelopment of the software ecosystem... |
| They committed to being self-sufficient... | ...indicating a broader institutional commitment to compute self-sufficiency. |
| It was hard to train because of memory... | ...memory constraints per device were identified as a primary bottleneck... |
The C2 Takeaway: Stop describing what is happening and start describing what the occurrence represents.