AITOPAI热报
ModelsMulti-source

美团开源 LongCat-2.5：1.6 万亿参数智能体模型，全程未用 Nvidia GPU 训练

Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU

Official accountDecoder@Jonathan Kemper
Blog/Media· 11 sources
Read the original
Why we picked it

美团把 1.6 万亿参数的 LongCat-2.5 权重直接放出来了，全程用华为昇腾芯片训练，Terminal Bench 拿了 88.4 分，只比 Claude Opus 5 低一点。想知道不用 Nvidia 能不能训出顶级智能体模型的，看看这套真实方案。

美团发布并开源了 LongCat-2.5 的模型权重，这是一款专为 AI 智能体设计的大模型，总参数量达 1.6 万亿，训练全程只使用中国国产算力，未提及任何 Nvidia GPU。在智能体任务基准 Terminal Bench 上，LongCat-2.5 取得 88.4 分，仅落后于 Claude Opus 5。

Body

在部署层面，示例方案运行在 194 颗华为昇腾（Ascend）芯片上，说明该模型从训练到推理都可以完全脱离美系 GPU 生态。这对受出口管制影响、难以采购 Nvidia 高端芯片的中国团队来说，是一条可复制的路线：用国产芯片完成万亿参数模型的训练，同时保持智能体能力处于第一梯队。

从行业格局看，LongCat-2.5 把国产算力的上限推到了与 Claude Opus 5 同一基准量级附近，开发者可以直接下载权重自行部署验证。对于需要在国内合规环境下搭建智能体系统的企业，以及关注芯片替代路线的工程团队，这套完全基于华为昇腾的训练与部署方案具有直接的参考价值。

Original · Decoder

Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents and trained entirely on Chinese hardware. On Terminal Bench, it scores 88.4 points, landing just behind Claude Opus 5. The sample deployment runs on 194 Huawei Ascend chips, and Meituan makes no mention of Nvidia GPUs. The article Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU appeared first on The Decoder .

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include LongCat-2.5, Meituan, Huawei Ascend, Claude Opus 5.

Source attribution: curated from Decoder (@Jonathan Kemper). Cross-source corroboration: 11 independent sources have covered this development (including rohanpaul_ai, Jerry Liu, TestingCatalog), confirming industry-wide signal strength.

#LongCat-2.5#Meituan#Huawei Ascend#Claude Opus 5#agents

Keep reading

Recent stories with the same tag or section
ModelsBuildersSelected22:20
Meituan open-sources LongCat-2.5 with 1.6T parameters

The MIT licence with commercial rights makes LongCat-2.5 one of the most permissive trillion-parameter releases, directly comparable to DeepSeek and Qwen open weights for agent and coding workloads. The 4-bit expert quantisation and chip co-design lower hardware barriers for teams running on non-NVIDIA accelerators common across China and Southeast Asia. Developers building long-context multimodal agents should evaluate it against closed APIs, with weights on Hugging Face and an API platform already live.

#LongCat-2.5#Meituan#open-source models#multimodal
Event
IT之家Source
ModelsBuildersMulti-sourceSelected23:13
LongCat-2.5 builds browser game for $0.28 vs Qwen3.8-Max's $0.82

The pricing detail matters more than the demo: cached-input economics at $0.006 per 1M tokens can make long-horizon agent runs dramatically cheaper than headline per-token rates suggest. Developers building multi-step coding agents can compare LongCat-2.5's cache-heavy cost profile against Qwen3.8-Max on their own workloads, with both models available via longcat.ai. The 1M-token training sequences also point to better endurance on tasks that span dozens of agent turns.

#LongCat-2.5#Qwen3.8-Max#Meituan#open-source models
Event
elvis@omarsar0Source
ModelsBuildersOfficial09:51
Chinese contentUniPat AI 推出 SPU 基准，AI 智能体预测真实湿实验室结果最高仅得 0.486

UniPat AI 拿 4207 个真实实验结果当考题，13 种智能体配置最高才考 0.486 分，基本等于瞎猜。比起刷数学题的排行榜，这个基准更能看出 AI 做科研的真实短板，做 AI4Science 的开发者应该跑一下自己的模型看看。

#UniPat AI#SPU benchmark#AI Agent#AI for Science
Primary sourcepandaily@[email protected] (Pandaily)Source
ModelsUS vs ChinaMulti-source01:15
Chinese contentGemini 4 Carbon 登顶 MathArena 数学榜，成本是 GPT-6.1 Sol 的 15 倍

数学榜换老大了：Google 的 Gemini 4 Carbon 干掉了 GPT-6.1 Sol，Anthropic 的 Fable 5.2 也挤进前排。但它的价格是 GPT-6.1 Sol 的 15 倍，做数学推理应用的朋友选型前得算算账。

#Gemini 4 Carbon#GPT-6.1 Sol#Fable 5.2#MathArena
Event
Official accountDecoder@Matthias BastianSource
ModelsBuildersSelected00:05
Perplexity releases Decider v1.3 and Decider Flash v1.3

The Apache 2.0 license plus per-task pricing of US$68 and US$34 puts both models among the cheapest options at their quality tier, a comparison point against larger closed models for routing and tool-calling pipelines. Teams building agent decision layers or extraction workflows should benchmark them against their current router models via Perplexity's API. The open weights also allow self-hosting for cost-sensitive deployments.

#Perplexity#Decider#open-source models#decision models
Official accountArtificial Analysis@ArtificialAnlysSource
ModelsBuildersSelected00:05
SubQ Snap 1.2 Gains 5 Points on Decision Index

The score jump is largely abstention calibration, not raw accuracy — a distinction buyers of decision or routing models should examine before benchmark comparisons. Teams deploying models for document triage or judgment tasks will care about the 62% all-pass rate on AA-Paperclip and the unchanged US$0.04/1M input pricing with a 384K context window. Compared with Drex Max v1.1 and GLiDE v1.2, the three leaders are effectively indistinguishable on the v2.1 index, so pricing and context limits become the practical differentiators.

#SubQ Snap#Subquadratic#Artificial Analysis#decision models
Official accountArtificial Analysis@ArtificialAnlysSource