美团开源 LongCat-2.0 并称全国产算力栈完成训练与适配，但未披露可量化订单或收入增量

中文網

繁體

简体

SUBSCRIBE

My account

Sign out

SUBSCRIBE

Trending Anthropic
HBM
AI Search
NTN Convergence
Inference Era AI
Advanced Packaging

Trending

SUBSCRIBE

CONNECT WITH US

Sign out

REALTIME NEWS

Global PMX revenue rises as advanced foundry and AI cooling demand strengthen

ICT

3min ago

Tesla starts public road tests of Cybercab

Electric Vehicles

3min ago

Microloops revenue hits record on cloud server cooling demand

ICT

4min ago

Sharp taps Foxconn for US$1.9 billion new-business push

ICT

4min ago

Meituan open-sources LongCat-2.0 as China's domestic AI stack gains momentum

ICT

5min ago

Satellite sovereignty goes global: Asia and Europe move to cut the Starlink cord

Aerospace

5min ago

Pegatron chairman downplays Intel CPU price hike worries

ICT

6min ago

Genius Electronic Optical June sales rise as mobile phone price impact appears limited

Communications

6min ago

ASE expands globally as Hanmi Semiconductor targets CoWoS packaging demand

Tomorrow's Headlines

Jul 7, 18:18

Kingsemi wins first commercial order for large-format panel-level packaging lithography tool

Tomorrow's Headlines

Jul 7, 18:18

Dongfeng Nissan seeks a turnaround amid weak Chinese auto market

Tomorrow's Headlines

Jul 7, 18:18

Apple's entry into foldable market boost supply chain orders despite cost pressure

Tomorrow's Headlines

Jul 7, 18:18

DRAM giants face US price-fixing lawsuit over HBM shift

Tomorrow's Headlines

Jul 7, 18:18

LG Innotek and LX Semicon reportedly enter Apple iPad OLED DDI supply chain

Tomorrow's Headlines

Jul 7, 18:18

Home

Tech

ICT

Meituan open-sources LongCat-2.0 as China's domestic AI stack gains momentum

Levi Li
, DIGITIMES, Taipei

Jul 8, 2026, 08:01

0

Credit: Meituan

Meituan has open-sourced LongCat-2.0, a trillion-parameter AI model trained and deployed entirely on Chinese computing infrastructure, turning the once-anonymous Owl Alpha into a key test of China's domestic AI model and chip ecosystem.

Meituan said on July 6 it had released LongCat-2.0's model weights, inference engine and core technical documents. Huawei Ascend, Moore Threads, and MetaX completed inference adaptation the same day, according to Jiemian News .
From Owl Alpha to LongCat-2.0
LongCat-2.0 first reached overseas developers as Owl Alpha, an anonymous OpenRouter model launched in late April, according to NBTimes . Despite lacking a brand, launch event or corporate identity, it surged into the global top three by monthly call volume within two months.
Owl Alpha reportedly processed 10.1 trillion tokens a month and 559 billion tokens a day, with monthly growth of 242%. It ranked first, second, and third by usage in Hermes Agent, Claude Code, and OpenClaw, respectively, as developers in the US, India, Kenya, Germany, and Spain used it for coding, tool calling, agent workflows, and model orchestration.
Owl Alpha's anonymous run ended on June 27, when overseas bloggers traced it to Meituan. The company confirmed it as LongCat-2.0 on June 30, fuelling debate over how a food delivery platform had built an AI model already embedded in global developer workflows.
Developer feedback was mixed but telling. Users praised Owl Alpha's coding and agent performance, including Unreal Engine blueprint-to-Python conversion and high-volume Hermes Agent runs without failed requests. Others said it was capable but slow, pointing to inference constraints. Either way, LongCat-2.0 had entered real developer workflows before Meituan attached its name.
A 50,000-card AI test
LongCat-2.0 uses a MoE architecture with 1.6 trillion total parameters, about 48 billion activated parameters per token and native support for a 1 million-token context window. Pre-training exceeded 30 trillion tokens, while peak training used more than 50,000 domestic accelerator cards, Sina reported.
Meituan began training large models on domestic computing infrastructure in 2023, starting with thousand-card clusters before scaling to tens of thousands of cards. Over three years, the LongCat team solved operator adaptation, communication optimisation, and distributed stability issues, completing training and inference on a 50,000-card domestic cluster.
The achievement matters because Chinese AI developers face limits in both computing power and software maturity. Domestic accelerators generally have less memory per card than Nvidia H800-class GPUs, while Huawei Ascend's CANN ecosystem remains far younger than Nvidia's CUDA stack. Training a trillion-parameter model under those constraints required system-level optimisation, not just more cards.
MoE was central to that strategy. By activating about 48 billion parameters per token instead of the full 1.6 trillion during inference, LongCat-2.0 reduces compute pressure and helps offset memory and performance limits in domestic AI hardware.
LongCat-2.0 is not being positioned as the world's strongest model. NBTimes noted that it still trails GPT-5.5 and Claude Opus 4.8 in some benchmarks. Its significance lies elsewhere: proving that large-scale Chinese model training and deployment can work on a fully domestic computing stack under hardware and software constraints.

Compiled by DIGITIMES, powered by AI
Huawei Ascend, Moore Threads and MetaX adapt
The adaptation work showed how quickly Chinese chip vendors are aligning with domestic large models. Huawei Ascend said deploying a 1.6 trillion-parameter model on the 64GB-per-card Ascend A2 while supporting high-performance inference and million-token sequences was a major challenge. Yicai reported that Huawei's CANN inference team used prefill-decode disaggregation and deployment-specific optimisations.
Moore Threads and MetaX also completed LongCat-2.0 inference adaptation the same day. The coordinated rollout suggests China's "domestic chip plus domestic model" strategy is moving from isolated breakthroughs toward broader ecosystem alignment.
LongCat-2.0 is available on GitHub and Hugging Face under the MIT licence, with BF16 original weights, FP8 quantised weights and INT8 quantised versions for free commercial use.
The open-source release came one week after LongCat-2.0 was introduced. Meituan said community feedback showed LongCat-2.0-Preview approached Claude Opus 4.6 in agent tasks such as tool calling and complex instruction execution, though it still trails Claude Opus 4.8.
Meituan's AI logic
Meituan's AI push has accelerated over the past year. In September 2025, it released LongCat-Flash-Chat, followed by reasoning model LongCat-Flash-Thinking and video generation model LongCat-Video.
Angel investor and AI expert Guo Tao told Yicai that LongCat-2.0 upgrades LongCat-1.0 in three areas: a 1.6 trillion sparse-parameter architecture, full-stack domestic computing through a 50,000-card cluster, and stronger agent adaptation through a 1 million-token context window, code workflows and automated task loops.
Meituan sees AI as mandatory. At this year's shareholder meeting, founder Wang Xing said AI transformation is a "mandatory question," not an elective one. Companies must use AI in products and at work, he said, adding that agents' impact on workflows has exceeded expectations from two years ago.
Meituan has launched products, including the Xiaotuan AI assistant. CFO Chen Shaohui said the company wants to participate in frontier technology through investment. Meituan holds stakes in listed firms, including Li Auto and Zhipu AI, plus unlisted companies such as Unitree Robotics, with total book value possibly exceeding CNY65 billion (US$9.6 billion).

Meituan CFO Chen Shaohui. Credit: The Org
NBTimes described Meituan as one of China's largest non-cloud AI investors, with early or significant stakes in Zhipu AI, Moonshot AI, Unitree Robotics, Moore Threads and MetaX. The portfolio aligns with Meituan's AI stack: Zhipu AI for large models, Moonshot AI for long-context technology, Unitree Robotics for embodied intelligence, and Moore Threads and MetaX for domestic GPU computing.
For Meituan,