According to a Reuters report on June 30, Meituan has released an open-source AI model LongCat-2.0, with 1.6 trillion parameters, adopting a sparse mixture of experts (Sparse MoE) architecture, trained entirely using domestic ASIC supercluster chips, without using any NVIDIA GPU or CUDA software stack, and the model’s context window reaches 1 million tokens.
Technical Specifications and Target Application Scenarios of LongCat-2.0
LongCat-2.0 adopts a sparse mixture of experts (Sparse MoE) architecture, similar to the Mixtral solutions from DeepSeek and Mistral: the model does not activate all 1.6 trillion parameters simultaneously; instead, an internal router selects a group of specialized sub-models for each token, reducing inference costs compared to a dense model of the same scale. The main technical specifications and deployment limitations are as follows:
Parameter Scale: 1.6 trillion (sparse MoE architecture, not all parameters activated simultaneously)
Context Window: 1 million tokens (DeepSeek-R1-0528 and GPT-OSS both have 128,000 tokens)
Training Hardware: Domestic ASIC supercluster (no NVIDIA GPU, no CUDA software stack)
Target Applications: AI agents, coding tools (code understanding, full-repository editing, automation tasks)
Deployment Form: Datacenter-level inference cluster, not supporting consumer-grade devices or most local deployments
Market Background of Domestic Chip Training and Bernstein Data
Meituan claims that the core inference architecture of LongCat-2.0 is portable and can run on existing hardware in China. This release comes as U.S. export controls continue to restrict the export of advanced AI chips to Chinese companies.
Equity research firm Bernstein estimates that NVIDIA currently holds about 40% of China's AI chip market share, with Huawei close behind; Bernstein also predicts that Huawei will make progress this year, causing NVIDIA's share in the Chinese market to drop by about 8 percentage points.
Current Status of Performance Claims: No Third-Party Verification Yet
Meituan has compared LongCat-2.0 with several closed-source models in published benchmark tests, but relevant reports indicate that these performance claims have not yet been independently verified by a third party.
The report also notes that optimization for domestic chips may limit LongCat-2.0's performance on NVIDIA hardware, which still dominates global data centers. Meituan says its core inference architecture remains portable, and independent tests will determine adoption willingness among developers outside China.
FAQ
What application scenarios are relevant for LongCat-2.0's 1 million token context window?
As of the report, DeepSeek-R1-0528 and OpenAI GPT-OSS both have a maximum context window of 128,000 tokens; LongCat-2.0 claims a 1 million token context window, which has potential significance for AI agent applications that need to process ultra-long codebases and complex task chains. However, these specification claims await independent verification.
What is Meituan's AI R&D background?
Meituan's core business is food delivery and local lifestyle services. It entered the AI space in 2023 by acquiring AI startup Light Year Beyond for $281 million, and only publicly announced its internal model plans in 2025. LongCat-2.0 is positioned as the inference engine for the company's AI agents and coding tools.
What are the advantages and disadvantages of LongCat-2.0's sparse MoE architecture compared to a 1.6 trillion dense model?
The core advantage of sparse MoE is that not all parameters are activated; routing to specific sub-models reduces inference computation costs compared to a dense model of the same scale. However, an architecture optimized for specific hardware (e.g., domestic ASIC) may have performance limitations on other hardware (e.g., NVIDIA GPU), and independent test results have not yet been released.
