Overview
LongCat-2.0 is a new-generation trillion-parameter MoE foundational large model officially open-sourced by Meituan on 2026/6/30 — the world's first trillion-level large model to complete the entire pre-training and inference R&D process entirely relying on a domestic 50,000-card computing cluster, requiring no overseas chips from start to finish. This is regarded by the industry as a milestone event where the domestic computing ecosystem truly runs through a full-stack closed loop.
The model has 1.6 trillion total parameters (MoE architecture, with an average activation of about 48B, activation range 33B~56B), employing a zero-computation expert mechanism for token-level dynamic activation — simple tokens consume no computing power, while complex tokens dispatch more experts to participate. It natively supports million-character ultra-long context, with a focus on optimizing Coding Agent capabilities — this path aligns with DeepSeek V4's Harness / Code Agent strategy, signaling that domestic large models have entered the 'Agent-first' phase.
Key Milestones:
- 2026/4/24: LongCat-2.0-Preview first released for open testing (released on the same day as DeepSeek V4, known in the industry as the 'twin stars' event for domestic large models)
- 2026/6/30: Official release, with weights, inference code, and fine-tuning tutorials fully released
- 2026/7/5: Fully open-sourced under the MIT commercial-friendly open-source license, becoming one of the first domestic trillion-level MoE large models to adopt the MIT license (more permissive than Apache 2.0, with fewer commercial restrictions)
- 2026/7/12: Simultaneously released domestic chip inference code (inference adaptation code for Huawei Ascend, T-Head, Cambricon, and other platforms all made public), extending the 'domestic computing full-stack closed loop' from Meituan's internal business to the entire community for reuse
In terms of positioning, LongCat deeply serves Meituan's local life services business (high-concurrency scenarios like in-store dining, food delivery, and rider dispatch), trained on Meituan's massive local life data, excelling in scenarios such as order parsing, user demand prediction, and local information generation. It is also available for public trial via the longcat.ai official website and has been launched on the OpenRouter global calling platform.
Key Features
- 1.6 Trillion MoE Architecture: Total parameters of 1.6T, placing it among the world's top-tier large models (on par with DeepSeek V4-Pro)
- World's First Domestic 50,000-Card Training: The first trillion-level model to complete the entire pre-training and inference process entirely relying on a domestic 50,000-card computing cluster, requiring no overseas chips
- Million-Character Ultra-Long Context: Native 1M Token context, capable of processing million-character industry documents at once
- Coding Agent Focused Optimization: Coding Agent is the core upgrade direction of the 6/30 version, aligning with the domestic Harness strategy
- Full Open Source · Free for Commercial Use: Weights, inference code, and fine-tuning tutorials fully released with no usage barriers
- Native Local Life Scenarios: Trained on Meituan's real business data: excels in order parsing, demand prediction, and local information generation
- Deep Optimization for Chinese Scenarios: Deeply fine-tuned for Chinese language data and local scenarios, outperforming overseas models of similar scale on Chinese tasks
Use Cases
- Long document / long conversation analysis (million-character context)
- Coding agent / programming Agent backend selection
- AI integration for local life applications (both within and outside Meituan ecosystem)
- Enterprises with compliance requirements for domestic computing deployment (state-owned enterprises / Xinchuang scenarios)
- Small and medium-sized enterprises and developers seeking free commercial use of trillion-parameter models
- Teams needing top-tier foundational model capabilities in Chinese scenarios
Pros
- 1.6T parameters + 1M context, same specifications as DeepSeek V4-Pro
- World's first domestic 50,000-card trained trillion-level model, benchmark for full-stack domestic production
- Full open source + free commercial use, no licensing costs
- Focus on optimizing Coding Agent, hitting the hottest track of 2026
- Backed by Meituan's real local life scenarios, with thorough engineering validation
- Forms a domestic open-source trio with DeepSeek V4 / Qwen
Pricing
Completely Free: From 6/30, model weights, inference code, and fine-tuning tutorials are all open-sourced, allowing free commercial use. The longcat.ai official website still provides free online experience. Enterprise-level private deployment is supported on demand by Meituan's technical team.
Summary
LongCat-2.0 is a landmark event for the full-stack domestic production of domestic large models in 2026 — the combination of 1.6T parameters + 50,000-card domestic cluster + full open source + free commercial use + Coding Agent optimization instantly places it in the top tier of open-source models. If you need a pure domestic stack for enterprise AI or are looking for the most cost-effective open-source base for Coding Agent, LongCat-2.0 is a must-test; for daily conversations and general tasks, it can still be used complementarily with Doubao / Kimi / DeepSeek.
Version History
- Meituan releases LongCat-2.5-Preview with a 1.6T MoE architecture and native image understanding (2026-09-25): Meituan launched LongCat-2.5-Preview with a 1.6T MoE architecture and 48B active parameters, a 1M context window and up to 128K output, adding native image understanding and compatibility with both the OpenAI and Anthropic APIs. Introductory pricing runs $0.30 per million input tokens, $1.20 for output and $0.006 for cache, with 5 million free tokens for existing users. No benchmark results were published.
- LongCat-2.0 and CatPaw AI operations platform shown at CIFTIS (2026-09-09): Meituan showed LongCat-2.0 at CIFTIS 2026: a 1.6T-parameter MoE model with roughly 48B average active parameters, native 1M-token context and an MIT license, described as the first trillion-parameter open model trained and served end to end on a ten-thousand-GPU domestic compute cluster. The company also introduced CatPaw, a general AI agent platform with mobile and desktop clients and 24/7 cloud execution, packaging industry skills such as review diagnostics, product copy generation, marketing asset review, campaign planning and operations analytics.
- Domestic chip inference code released (2026-07-12): Inference adaptation code for domestic chip platforms such as Huawei Ascend, T-Head, and Cambricon all made public, extending the 'domestic computing full-stack closed loop' from Meituan internally to the entire community for reuse
- Full open source under MIT license (2026-07-05): Fully open-sourced under the MIT commercial-friendly open-source license, becoming one of the first domestic trillion-level MoE large models to adopt the MIT license (more permissive than Apache 2.0)
- LongCat-2.0 official release (2026-06-30): Weights, inference code, and fine-tuning tutorials fully released; world's first trillion-level model to complete full R&D entirely relying on a domestic 50,000-card cluster; zero-computation expert mechanism for token-level dynamic activation
- LongCat-2.0-Preview (2026-04-24): First released for open testing; trillion-level MoE + 1M context; released on the same day as DeepSeek V4, known as the 'twin stars' event for domestic large models