TECH INSIGHT • 深度智库
LongCat-2.0 发布,我模拟了 5 万卡训练集群,发现 6 个硬核亮点
作者:刀哥 | 聚焦 AI Infra 与核心硬核工程大家好,我是刀哥。做过大厂研发、做过出海硬件,现在深耕 AI 编程和 AI 工具。
7月1日,美团发布了 LongCat-2.0,一时间各种“国产大模型再突破”“代码能力挑战 Claude”的稿子铺天盖地。
宣传里有一个亮点我很感兴趣 “业界首个在五万卡国产算力集群上完成全流程训练与推理的万亿参数模型”(美团官方并未点名具体型号。数智前线获悉,LongCat-2.0在昇腾5万卡集群上完成训练和推理全流程。)
于是我购买了token资源包,用LongCat-2.0做了MegaTrainSim 的项目——一个模拟 5 万卡 GPU 训练集群调度的仿真系统。动手做这个项目,初衷就是希望摸透一个大模型究竟是怎么被“炼”出来的,以及要把它的工程做扎实,到底难在哪儿,也测试下longcat的编程能力。
用这个仿真视角重新审视 LongCat-2.0,很多隐藏在水面之下的核心细节反而清晰地暴露了出来。今天就跟大家聊聊:LongCat-2.0 的技术含金量究竟在哪里?它的实际模型能力,到底是接近 Sonnet 4.6,还是有底气叫板 Opus 4.8?
💡 核心结论
LongCat-2.0 的 Infra(基础设施)成就远大于其模型科学层面的突破。 它最硬核的成绩,是成功在国产 AI ASIC 五万卡集群上,将预训练、推理、稳定性、算子、通信及框架的全链路彻底跑通。这种底层工程的含金量极高,我们逐一拆解。
亮点一:五万卡国产算力闭环的底层真相
这是最硬核的一条,也是目前外界解析得最含糊的一条。多数报道只停留在“五万张国产卡”的数字繁荣上,却选择性忽略了它在物理拓扑上的恐怖复杂度。
在构建 MegaTrainSim 仿真模型时,光是精确建模这五万卡集群的物理拓扑就让我吃尽了苦头。结合国内液冷高密机房的实际承载限制,一个万卡级集群的真实物理结构需要呈现出高度展开的层级拓扑:
⚡ 5万卡集群物理级联架构 (Cluster Topology)
集群层➔包含 10 个独立 Pod 机房核心单元Pod 层➔每个 Pod 部署 100 个现代标准液冷机柜 (Rack)机柜层➔每个 机柜 科学容纳 6 个高密计算节点 (Node)节点层➔每个 节点 挂载 8 张国产核心 GPU/NPU单机柜功耗维持在安全的 ~50kW 液冷红线,总物理卡数 = 10 × 100 × 6 × 8 = 48,000 张(外加容灾冗余节点凑齐五万)
(ps:NVIDIA SuperPOD 或美系高密度机房一个机柜可以做到 12个node)
在这种恐怖的体量下,要保证高效率,必须完美驾驭四层完全不同的通信带宽与通信延迟:
| 通信物理范围 | 示例带宽 | 典型计算场景 |
|---|---|---|
| 同节点内 (NVLink 级别) | 900 GB/s | TP (张量并行) 组内 AllReduce |
| 同机柜跨节点 | 400 GB/s | EP (专家并行) 组 All-to-All |
| 同 Pod 跨机柜 | 200 GB/s | DP (数据并行) 梯度同步 |
| 跨 Pod 全局通信 | 100 GB/s | 最昂贵的底层网络吞吐瓶颈 |
这里有必要揭开国产 ASIC 真正的技术痛点。 NVIDIA 体系能够凭借 NCCL、NVLink、NVSwitch 以及 InfiniBand 构筑稳固的生态,经历了长达数年的社区踩坑和 debug 工具链沉淀。而国产芯片的通信库、算子库和编译器的成熟度,起点天然更加粗糙。
美团技术报告里重点提及的“通信异常处理、弹性扩缩卡、自动故障恢复、确定性算子、Bitwise 一致性验证”,绝非锦上添花的附加功能。这代表着工程师在一个更脆、更复杂的底层基础设施上,用软件和架构生生将底座补齐。这比在成熟的 H100 集群上做同样的优化,硬核了一个数量级。
亮点二:万亿 MoE 的调度才是真正的吞吐难题
LongCat-2.0 拥有 1.6T 总参数,平均激活约 48B,是一个典型的超大型 MoE 模型。许多人一看到 1.6T 就会下意识代入超大 Dense 模型,这种认知是不对的。MoE 的绝大多数参数在计算时处于“低能耗”的冷态——每个 Token 仅路由激活极少部分的专家,实际计算开销更贴近几十 B 的 dense 模型。
然而,这种设计带来了调度上的致命死穴:All-to-All 全局网络通信。
📊 MegaTrainSim 仿真:LongCat 规模 MoE 单步通信开销
MoE 专家并行 All-to-All (最昂贵)350 GB / stepTP 组内 AllReduce120 GB / stepPP 流水线 P2P 通信80 GB / step在 MoE 模型里,每个 Token 都必须被高频路由给特定的专家进行计算,而这些专家散落驻留在不同的物理卡上。这意味着每向前训练推进一步,集群就要进行一次高压的 All-to-All 数据分发与回收。一旦路由发生倾斜,某些“热门专家”就会产生严重的过载热点。负责这几张卡的 GPU 将满载卡死,而其余大量 GPU 只能陷入空等状态,极大地拖慢整个集群的 Step time。
在我的 MegaTrainSim 仿真实验中,我模拟了同样万亿规模的 MoE 作业,结果发现在 4096 卡的基准下,如果位置放置(Placement)策略不合理,All-to-All 的通信耗时就会被拉开 3倍以上 的恐怖代差。
管中窥豹,由此在工程上倒推美团的处境: 4096 卡尚且如此,美团在 5 万卡这种物理拓扑更深、网络跳数(Hop)更多的国产算力集群上,不仅要解决路由倾斜,还要压制这种指数级放大的 All-to-All 通信气泡。他们能把全链路死死压制在安全线之内,技术团队这次确实啃下了行业内最硬的一块硬骨头。
亮点三:长上下文稀疏 Attention 的精细剪裁
LongCat-2.0 主打 1M 上下文,底层依托其定制的 LongCat Sparse Attention。长序列训练的永恒矛盾在于:传统 Attention 的计算复杂度呈现高昂的 O(n²) 爆发。当序列长度翻倍时,算力开销将面临四倍的暴击,使得 1M 的 Full Attention 在主流集群里几乎没有实操可能。
稀疏 Attention 斩断了这一死结。它的核心逻辑放下了“每个 Token 都必须时刻与其余所有 Token 建立注意力关联”的执念。
在真实的工程落地中,长代码文件里的某段核心逻辑,通常只强烈依赖于前面十多行处的变量定义或特定的类声明,跨文本扫描整座代码库里的每个字符纯属浪费。虽然 Gemini 的长上下文路线或 Qwen 的滑动窗口在做相似的策略,但 LongCat 能够将其稳定地沉淀到代码库级别的 Agent 生产任务中,其针对工程研发场景的落地价值非常直接。
亮点四:动态计算——“零计算专家”的极致剪枝
在 LongCat-2.0 披露的技术栈中,ScMoE、MOPD 多专家融合以及“零计算专家”的概念非常出彩。传统的动态路由多多少少会留存一些冗余的传输开销,而“零计算专家”机制彻底走向了精细化算力调配。
只要路由器在底层判断出某个传入的 Token 在当前语义下无需调用特定专家,该节点就会被连根拔除——直接阻断其计算,并且连带抹去该 Token 的跨卡通信过程,从源头上豁免了 All-to-All 整体网络的带宽承载压力。
这项优化正在顺应如今大模型能效比演进的主流洪流:不同的 Token 与不同权重的任务,天然应该去消耗完全不对等的动态计算资源。宁可把算力花在极致的逻辑对撞上,也不让毫无意义的空转 Token 白白烧掉昂贵的机房带宽。
亮点五:弹性故障恢复——重新定义生产稳定工程
在五万卡的惊人规模下跑万亿级模型训练,故障永远不是偶发的意外,它是不容辩驳的日常常态。哪怕强如 Meta,在发布 Llama 3.1 405B 的官方技术报告时也得坦白:在操盘 16K H100 集群进行千亿参数训练期间,出现了极其密集的硬件中断事故,平均每隔几个小时就会遭遇一次意料之外的 Unexpected Shutdown。那么当卡数堆叠到五万卡级别时,这个崩溃频率只会呈指数级上升。
在 MegaTrainSim 的仿真模拟中,我对集群里的高频灾难进行了四类归化建模:
⚠️ GPU_THROTTLE (硬件降频)常态隐形杀手GPU 局部过热引发过载降频。卡没有死,但算力极度挂车,拖慢全局 Rank 步调。
⚠️ NIC_ERROR (网卡瞬时异常)高频抖动源网络物理层抖动导致的短暂失联。处理不好极易触发通信库全局锁死挂起。
⚠️ SWITCH_CONGESTION (交换机拥塞)范围性冲击数据突发吞吐限死,波及整座机柜 (Rack),造成局部流水线气泡无限放大。
⚠️ NODE_CRASH (物理节点宕机)毁灭性硬中断物理机直接挂死掉线。整个 Job 必须强制熔断,并依赖上一个 Checkpoint 回滚。
通过泊松过程在仿真器中灌入灾难事件流,我跑出了一项反直觉的底层数据:吞吐效率在 p99 的核心跌落,其罪魁祸首并非毁灭性的物理机宕机(NODE_CRASH),而是半死不活的 GPU 降频和网卡偶发闪断。 宕机会立刻引起系统重启和任务回滚,边界非常清晰;而局部降频则属于“维持半条命”,它在死死拖垮整组并行环路的木桶短板。LongCat 重点突破的“慢 Rank 定位、坏卡弹性隔离、自动化免中断恢复”,正是死磕这一顽疾的极佳良药。
亮点六:MFU(模型浮点利用率)暴涨 1.5 倍的工程内幕
美团官方表示,在一连串密集的 Infra 优化技术栈注入后,其训练长序列时的 MFU(模型浮点利用率)迎来了 1.5 倍的巨大提升。
要明确的是,MFU 是刺破一切营销话术、评估集群训练效率的核心死指标。它直接刻画了“你花真金白银买过来的物理算力中,到底有多少比例在真刀真枪地执行有效计算”。传统的千卡或万卡 H100 集群,在工程团队打满补丁的情况下,MFU 通常徘徊在 35%–55% 之间。而国产 AI 芯片由于历史原因,早期常面临 MFU 低下的窘境,主要原因在于三大死结:
01Kernel 算子融合做得很薄 频繁的中间结果导致严重的显存访存开销,算力大量被卡在 I/O 等待中。
02网络通信与核心计算未能实现深度 Overlap 算力执行时,通信网络空转;网络传输数据时,GPU 停工等待。
03专用编译器与静态图优化不成熟 无法根据极其复杂的并行拓扑切分出帕累托最优的计算图执行路径。
因此,美团在千亿/万亿长序列极端场景下,相比未优化前的基线实现了 1.5 倍的相对效率飞跃。这说明了两件事:国产卡集群在早期研发时的原始优化空间极为广阔;美团的 AI Infra 团队确实下足了笨功夫,硬生生通过改写底层 kernel 把这些深坑全部填平了。
🛠️ 仿真背书:我用 MegaTrainSim 验证了什么?
在 MegaTrainSim 仿真系统里,我利用基于 Heap 的事件驱动引擎,高度复现了这套调度方案的逻辑质量。
其拓扑感知调度器(Topology-Aware Scheduler)彻底抛弃了盲目的 Best-Fit 模式。它严格遵循核心评分公式:score = α×locality - β×fragmentation - γ×congestion - δ×failure_risk。在仿真分发任务时,它会强行把 TP 组塞进同节点(享受 900 GB/s 顶配带宽),把 EP 组聚拢在同机柜,确保跨 Pod 机房单元通信的流量被死死卡在安全线下。配合泊松过程生成的 NODE_CRASH 灾难追踪回滚链路,这套仿真的逻辑自洽度极高。
LongCat-2.0 的模型能力到底几斤几两?
尽管美团给出的官方 Benchmark 截图里,LongCat-2.0 在 SWE-bench Pro 以及 Terminal-Bench 等前沿评测集上拿到了“接近甚至超越部分闭源头部大模型”的数据,但我们必须保持冷静客观。根据当前的公开资料,我个人的客观判断如下:
与 Claude Sonnet 4.6 相比: 差距真实存在,但这绝不是落后一代的“代差”。Anthropic 经历过数年极致 RLHF 以及 Constitutional AI 的语料清洗,其在多步复杂代码推理和工具链调用的高确定性上拥有惊人的护城河优势。LongCat-2.0 在这些极高难度的指标上实现了深度迫近,这本身已是极其强悍的成果,但“逼近”和“完全超越”不可混为一谈,目前缺乏大面积的第三方独立测算数据。
与 Claude Opus 4.8 相比: 这里的技术鸿沟依然在数个月的研发量级。Opus 4.8 在极长程代码重构与跨模块多轮工具调用的意图对齐上,展现了全球头部的智力水平。LongCat-2.0 想要跨越这道坎,绝不是单纯在国产集群上无脑堆大参数量就能解决的,背后还涉及到 Alignment 数据纯净度、强化学习 Pipeline 的调校精细度以及推理时计算(Inference-time Compute)核心策略的综合性差距。
| 核心考察评估维度 | LongCat-2.0 | Claude Sonnet 4.6 | Claude Opus 4.8 |
|---|---|---|---|
| 代码 Agent 实战实力 | 强(基于官方自测数据) | 极其强悍(全球工业级公认) | 全球统治级别(多轮一致性极高) |
| 原生输入上下文窗口 | 1M (稀疏 Sparse 动态剪裁) | 1M (全量密集高精度扩容) | 1M (最顶级全量注意力保持) |
| 大代码块最大输出上限 | 中等 (受限于专家的激活频率) | 高 (满足大批文件跨模块修改) | 128K (极限级复杂代码输出) |
| 模型能力的独立公测 | 待权重开源(社区目前尚在等待) | 已通过广泛开源工业洗礼 | 已通过广泛开源工业洗礼 |
| 企业级边际推理成本 | 单步激活 ~48B (硬件成本极低) | 性价比优异 (支持 Prompt Caching) | 超长上下文会话时成本高昂 |
| 国产芯片算力生态适配 | ✅ 5万卡全链路实打实跑通 | ❌ 完全无法适配国产 ASIC | ❌ 完全无法适配国产 ASIC |
美团 LongCat-2.0: 它是在算力底座更薄、更脆的国产 ASIC 集群上,依靠“Sparse Attention 稀疏裁切 + 零计算专家路由”把上下文逼上了 1M。
一个极其现实且重要的底层前提: 目前 LongCat-2.0 的物理模型权重尚未开放公开发载(GitHub 页面上高挂着 "Model weights coming soon")。这意味着,当前的傲人数据皆为官方自测,尚未经历社区黑盒测试的洗礼。等权重正式开源、真实复现之后,才是它迎战大风大浪的开始。
不过,它的战略生态位早已坚不可摧——它的最大历史使命并不是要在开卡智商上立刻碾压 Opus 4.8,而是雄辩地证明:在极限地缘政治生存环境下,我们完全可以用纯国产基础设施支撑起一条高转化率、低推理成本的代码万亿 MoE 全链路。这个现实意义,在国内产业界价值千金。
说在最后
如果没有亲手推敲 MegaTrainSim 这个集群仿真项目,我对模型大牛们的直觉往往流于表面。以前我认为堆砌大模型无非就是把大量算力卡攒到一块执行梯度更新。
现在我深刻明白:真正的天堑不在于有多少张卡,而是有没有能力突破万卡集群在物理通信的极限瓶颈。避免被坏卡无限拖累、避免被通信死锁扼杀咽喉,避免不合理的拓扑打散整体逻辑并行组,这需要真正的硬核功夫。
LongCat-2.0 的这场仗,明确告诉了所有人,美团绝不仅仅是一家局限于在应用层打转的本地生活公司。他们在国产芯片 Infra 层表现出的工程执念与踩坑沉淀,是有着极高门槛、值得整个行业认真研读的硬核样本。
关于真正的模型上限,等它开源,等全球开发者把它的 API 真正接入到真实的 IDE 生产项目里高压测试满一个月。到那个时候大家亮出来的口碑,才真正算数。
另外,我实名认证后loncat送了1000万token,还买了9.9的5000万token包,做了4个项目,结果发现,连送的token都用不完,不得不说,LongCat太耐用太实诚了!
📚 核心参考来源与文献储备
[1] 美团官方新闻发布:https://www.meituan.com/news/NN260630164005904
[2] LongCat-2.0 官方开源仓库:https://github.com/meituan-longcat/LongCat-2.0
[3] LongCat-Flash 底层技术报告:arXiv:2509.01322v1
[4] DeepSeek-V3 核心架构效率报告:arXiv:2412.19437v1
[5] Meta Llama 3.1 16K计算集群稳定性技术复盘 - Meta AI Blog
这篇笔记是我的第一篇AI infra学习笔记,我会更新一系列文章,完成整个AI Infra的学习,抛砖引玉,欢迎评论交流。
#LongCat#美团AI#大模型训练#AIInfra#MoE调度
TECH INSIGHT • Deep Insight
LongCat-2.0 released — I simulated a 50,000-GPU training cluster and found 6 hardcore highlights
Hi everyone, I'm Brother Dao. I've worked in R&D at a major tech company and built overseas hardware; now I focus deeply on AI coding and AI tools.
On July 1, Meituan released LongCat-2.0, and suddenly the internet was flooded with headlines like "Another domestic large model breakthrough" and "Coding capability challenges Claude."
One highlight in the marketing caught my interest: "The industry's first trillion-parameter model to complete the full training and inference pipeline on a 50,000-card domestic computing cluster" (Meituan's official announcement did not name a specific model. Shuzhi Qianxian learned that LongCat-2.0 completed the full training and inference pipeline on an Ascend 50,000-card cluster.)
So I purchased a token package and used LongCat-2.0 to build the MegaTrainSim project — a simulation system that models scheduling on a 50,000-GPU training cluster. My original motivation for this project was to understand exactly how a large model is "forged," where the real engineering difficulty lies, and also to test LongCat's coding capability.
💡 Core conclusion
LongCat-2.0's Infra achievements far outweigh its breakthroughs at the model-science level. Its most hardcore accomplishment is successfully running the entire pipeline — pretraining, inference, stability, operators, communication, and framework — on a domestic AI ASIC cluster of 50,000 cards. The value of this kind of low-level engineering is extremely high. Let's break it down one by one.
Highlight 1: The underlying truth of the closed loop of 50,000 domestic cards
This is the most hardcore point, and also the one outside observers have explained most vaguely. Most reports stop at the numerical glory of "50,000 domestic cards" and selectively ignore the terrifying complexity of its physical topology.
When building the MegaTrainSim simulation model, just accurately modeling the physical topology of this 50,000-card cluster gave me enormous trouble. Given the actual hosting constraints of domestic liquid-cooled high-density data centers, the real physical structure of a 10,000-card cluster needs to exhibit a highly expanded hierarchical topology:
⚡ 50,000-card cluster physical cascade architecture (Cluster Topology)
Single-rack power consumption stays at a safe liquid-cooling red line of ~50kW; total physical cards = 10 × 100 × 6 × 8 = 48,000 (plus disaster-recovery redundant nodes to make up 50,000)
(PS: An NVIDIA SuperPOD or a US-style high-density data center can fit 12 nodes per rack)
At this terrifying scale, achieving high efficiency requires perfectly managing four completely different layers of communication bandwidth and latency:
Here it's necessary to reveal the real technical pain point of domestic ASICs. The NVIDIA ecosystem was able to build a solid ecosystem with NCCL, NVLink, NVSwitch, and InfiniBand, backed by years of community debugging and debug toolchain accumulation. Domestic chips, by contrast, started with inherently rougher communication libraries, operator libraries, and compilers.
The "communication exception handling, elastic scaling of cards, automatic fault recovery, deterministic operators, and Bitwise consistency verification" emphasized in Meituan's technical report are by no means decorative add-ons. They represent engineers using software and architecture to fill in the foundation on a more fragile and more complex underlying infrastructure. This is an order of magnitude harder than doing the same optimization on a mature H100 cluster.
Highlight 2: Scheduling a trillion-parameter MoE is the real throughput challenge
LongCat-2.0 has 1.6T total parameters and an average of ~48B activated, making it a typical ultra-large MoE model. Many people instinctively treat 1.6T as an ultra-large dense model, but that understanding is wrong. In MoE, the vast majority of parameters are in a "low-power" cold state during computation — each token routes to and activates only a small fraction of experts, so the actual compute cost is closer to a dense model of tens of billions of parameters.
However, this design introduces a fatal scheduling weakness: All-to-All global network communication.
📊 MegaTrainSim simulation: single-step communication overhead for a LongCat-scale MoE
In an MoE model, every token must be frequently routed to specific experts for computation, and these experts are scattered across different physical cards. This means that for every training step forward, the cluster must perform a high-pressure All-to-All data distribution and collection. Once routing becomes skewed, certain "hot experts" create severe overload hotspots. The GPUs responsible for those cards run at full capacity and stall, while a large number of other GPUs are left idle waiting, dramatically slowing down the cluster's step time.
In my MegaTrainSim simulation experiments, I simulated MoE jobs of the same trillion-parameter scale and found that at a baseline of 4096 cards, if the placement strategy was unreasonable, All-to-All communication time could diverge by a terrifying factor of more than 3x.
Looking at the bigger picture through this small window, we can infer Meituan's engineering situation: if even 4096 cards behaves like this, then on a 50,000-card domestic computing cluster with deeper physical topology and more network hops, Meituan had to not only solve routing skew but also suppress this exponentially amplified All-to-All communication bubble. The fact that they managed to keep the entire pipeline tightly within the safety line means the technical team really chewed through one of the hardest bones in the industry.
Highlight 3: Fine-grained tailoring of long-context sparse attention
LongCat-2.0 focuses on 1M context, built on its custom LongCat Sparse Attention. The eternal contradiction of long-sequence training is that traditional Attention has explosive O(n²) computational complexity. When sequence length doubles, compute cost faces a 4x hit, making 1M Full Attention practically impossible on mainstream clusters.
Sparse Attention cuts through this deadlock. Its core logic lets go of the obsession that "every token must maintain attention connections with all other tokens at all times."
In real engineering deployment, a core piece of logic in a long code file usually depends strongly only on variable definitions or specific class declarations from a dozen or so lines earlier; scanning every character across the entire codebase is pure waste. Although Gemini's long-context route and Qwen's sliding window do something similar, LongCat has stably embedded this into agent production tasks at the codebase level, giving it very direct value for engineering R&D scenarios.
Highlight 4: Dynamic computation — extreme pruning with "zero-compute experts"
In LongCat-2.0's disclosed technology stack, the concepts of ScMoE, MOPD multi-expert fusion, and "zero-compute experts" stand out. Traditional dynamic routing tends to retain some redundant transmission overhead, but the "zero-compute expert" mechanism pushes compute allocation to an extreme level of precision.
As long as the router determines at the bottom layer that an incoming token does not need a specific expert under the current semantics, that node is removed at the root — its computation is directly blocked, and the token's cross-card communication process is also eliminated, exempting the overall All-to-All network from bandwidth pressure at the source.
This optimization aligns with the mainstream trend in large-model energy-efficiency evolution: different tokens and tasks with different weights should naturally consume completely asymmetric dynamic compute resources. Better to spend compute on extreme logical collision than to let meaningless idle tokens burn expensive data-center bandwidth.
Highlight 5: Elastic fault recovery — redefining production stability engineering
At the astonishing scale of 50,000 cards running trillion-parameter model training, failures are never rare accidents; they are an undeniable daily reality. Even Meta, in its official technical report for Llama 3.1 405B, had to admit that during training on a 16K H100 cluster with hundreds of billions of parameters, hardware interruptions were extremely dense, with an unexpected shutdown occurring on average every few hours. When you stack the card count to 50,000, that crash frequency only rises exponentially.
In the MegaTrainSim simulation, I modeled the high-frequency disasters in the cluster into four categories:
GPU local overheating triggers overload-induced frequency throttling. The card isn't dead, but compute is severely degraded, slowing down the global rank step.
Transient disconnection caused by physical-layer network jitter. If not handled well, it can easily trigger a global lockup and hang in the communication library.
Sudden data throughput limits affect an entire rack, causing local pipeline bubbles to expand without bound.
A physical machine dies outright. The entire job must be forcibly interrupted and rolled back to the last checkpoint.
By injecting disaster event streams into the simulator via a Poisson process, I obtained a counterintuitive underlying finding: the culprit behind the p99 throughput collapse was not catastrophic physical machine crashes (NODE_CRASH), but half-dead GPU frequency throttling and occasional NIC link flaps. A crash immediately triggers system restart and job rollback, with very clear boundaries; local throttling, by contrast, "keeps half a life" and acts as the short plank dragging down the entire parallel loop. LongCat's focused breakthroughs in "slow rank detection, elastic isolation of bad cards, and automatic interruption-free recovery" are precisely the right medicine for this persistent disease.
Highlight 6: The engineering story behind MFU (model FLOPs utilization) surging 1.5x
Meituan officially stated that after a dense stack of Infra optimizations, its MFU (model FLOPs utilization) during long-sequence training improved by a massive 1.5x.
To be clear, MFU is the core hard metric that pierces through all marketing talk and evaluates cluster training efficiency. It directly measures "of the physical compute you paid real money for, how much is actually executing useful computation." Traditional H100 clusters of thousands or tens of thousands of cards, with engineering teams applying all patches, typically hover at 35%–55% MFU. Domestic AI chips, for historical reasons, often suffered from low MFU early on, mainly due to three deadlocks:
Kernel operator fusion was too shallow. Frequent intermediate results caused severe memory access overhead, leaving much of the compute stuck waiting on I/O.
Network communication and core computation failed to overlap deeply. While compute was executing, the communication network idled; while the network transmitted data, the GPUs stopped and waited.
Dedicated compilers and static graph optimization were immature. They could not split the compute graph into a Pareto-optimal execution path based on extremely complex parallel topology.
Therefore, in the extreme long-sequence scenario of hundreds of billions or trillions of parameters, Meituan achieved a 1.5x relative efficiency improvement over the unoptimized baseline. This shows two things: the original optimization headroom for domestic card clusters in early R&D was extremely large; and Meituan's AI Infra team truly put in the hard work, filling in all those deep pits by rewriting low-level kernels.
🛠️ Simulation endorsement: What did I verify with MegaTrainSim?
In the MegaTrainSim simulation system, I used a Heap-based event-driven engine to reproduce the logical quality of this scheduling scheme with high fidelity.
Its Topology-Aware Scheduler completely abandons blind Best-Fit mode. It strictly follows the core scoring formula: score = α×locality - β×fragmentation - γ×congestion - δ×failure_risk. When dispatching tasks in simulation, it forcibly places TP groups on the same node (enjoying 900 GB/s top-tier bandwidth) and clusters EP groups within the same rack, ensuring that cross-Pod data-center-unit traffic stays tightly under the safety line. Combined with the NODE_CRASH disaster-tracking rollback chain generated by the Poisson process, the simulation has extremely high logical consistency.
What is LongCat-2.0's actual model capability?
Although the official benchmark screenshots provided by Meituan show LongCat-2.0 achieving "close to or even exceeding some closed-source top models" on frontier benchmarks such as SWE-bench Pro and Terminal-Bench, we must remain calm and objective. Based on currently available public information, my personal objective assessment is as follows:
Compared with Claude Sonnet 4.6: The gap is real, but it is by no means a "generational gap" of falling behind by one generation. Anthropic has gone through years of extreme RLHF and Constitutional AI corpus cleaning, giving it a formidable moat in high-determinism multi-step complex code reasoning and toolchain invocation. LongCat-2.0 has narrowed the gap deeply on these extremely difficult metrics, which in itself is an extremely strong achievement, but "approaching" and "fully surpassing" must not be conflated, and there is currently a lack of large-scale third-party independent evaluation data.
Compared with Claude Opus 4.8: The technical gap here remains on the order of several months of R&D. Opus 4.8 demonstrates world-leading intelligence in extremely long-range code refactoring and cross-module multi-turn tool-call intent alignment. For LongCat-2.0 to cross this threshold, it cannot be solved simply by piling up large parameter counts on domestic clusters; behind it lies a comprehensive gap involving alignment data purity, the precision of reinforcement learning pipeline tuning, and core strategies for inference-time compute.
Meituan LongCat-2.0: It pushed context to 1M on a thinner, more fragile domestic ASIC cluster by relying on "Sparse Attention pruning + zero-compute expert routing."
An extremely realistic and important underlying premise: LongCat-2.0's physical model weights have not yet been publicly released for download (the GitHub page prominently displays "Model weights coming soon"). This means the current impressive numbers are all official self-test results that have not yet undergone community black-box testing. Only after the weights are officially open-sourced and truly reproduced will it be the start of facing the big waves.
However, its strategic ecological niche is already unshakable — its greatest historical mission is not to immediately crush Opus 4.8 in raw intelligence, but to convincingly prove that under extreme geopolitical survival conditions, we can fully use domestic infrastructure to support a complete trillion-parameter MoE pipeline for code with high conversion efficiency and low inference cost. This practical significance is worth a fortune to the domestic industry.
Final words
If I had not personally worked through the MegaTrainSim cluster simulation project, my intuition about model experts would have stayed superficial. Previously, I thought building large models was nothing more than piling up a large number of compute cards and running gradient updates together.
Now I deeply understand: the real chasm is not how many cards you have, but whether you can break through the physical communication bottleneck of a 10,000-card cluster. Avoiding being endlessly dragged down by bad cards, avoiding being strangled by communication deadlocks, and avoiding unreasonable topology that breaks apart logical parallel groups — this requires true hardcore skill.
LongCat-2.0's battle clearly tells everyone that Meituan is far more than a local-life company stuck at the application layer. Their engineering obsession and accumulated lessons at the domestic chip Infra layer form a hardcore sample with extremely high barriers that the entire industry should study carefully.
As for the real ceiling of the model — wait until it's open-sourced, and wait until developers around the world connect its API into real IDE production projects and stress-test it for a full month. Only the reputation revealed by everyone at that point truly counts.
Also, after real-name verification, LongCat gave me 10 million free tokens, and I also bought the 9.9-yuan package of 50 million tokens and built 4 projects. It turned out I couldn't even use up the free tokens. I have to say, LongCat is incredibly durable and honest!
📚 Core references and literature reserve
[1] Meituan official press release: https://www.meituan.com/news/NN260630164005904
[2] LongCat-2.0 official open-source repository: https://github.com/meituan-longcat/LongCat-2.0
[3] LongCat-Flash underlying technical report: arXiv:2509.01322v1
[4] DeepSeek-V3 core architecture efficiency report: arXiv:2412.19437v1
[5] Meta Llama 3.1 16K compute cluster stability technical review - Meta AI Blog
This note is my first AI infrastructure learning note. I will publish a series of articles to complete my study of AI Infra. This is my humble attempt to spark discussion — comments and exchanges are welcome.
#LongCat #MeituanAI #LargeModelTraining #AIInfra #MoEScheduling