Meituan Reveals LongCat-2.0 on 30 June 2026
On 30 June 2026, Meituan — China's dominant food delivery and e-commerce platform — open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts large language model trained entirely on domestic Chinese AI chips. The release came with model weights published on GitHub and Hugging Face under an MIT licence, full API access through OpenAI- and Anthropic-compatible endpoints, and an immediately competitive price point. What made the release genuinely significant, however, was the story that preceded it: for several weeks before the official announcement, LongCat-2.0 had been quietly topping OpenRouter's global usage charts under the anonymous alias "Owl Alpha."
Architecture: 1.6 Trillion Parameters, 48 Billion Active
LongCat-2.0 uses a Mixture-of-Experts design that decouples total parameter count from inference cost. Of its 1.6 trillion total parameters, only around 48 billion are activated for any given token, making per-request compute comparable to a dense model a fraction of its total size. The model ships with a native 1-million-token context window, maintained by a custom linear-complexity attention mechanism the team calls LongCat Sparse Attention, which prevents memory cost from scaling quadratically at extreme context lengths. The architecture is organised into three expert groups — Agent, Reasoning, and Interaction — designed to handle agentic multi-step tasks, complex chain-of-thought reasoning, and conversational interaction respectively.
The Owl Alpha Story: Topping OpenRouter in Stealth
Before the 30 June announcement, LongCat-2.0 was accessible on OpenRouter under the codename "Owl Alpha." During this undisclosed period, it accumulated substantial usage from developers who noted unusually strong performance on coding and agentic tasks — performance that placed it consistently at the top of OpenRouter's usage rankings. Meituan's decision to run a stealth deployment before the public announcement meant the model accumulated real developer feedback and production usage data before the reveal, a different approach from the closed-access beta that most frontier AI labs operate.
Benchmarks: Edging Past GPT-5.5 on Coding Tasks
On SWE-bench Pro — an industry-standard evaluation of a model's ability to resolve real GitHub software engineering issues — LongCat-2.0 scored 59.5, edging past the 58.6 reported for OpenAI's GPT-5.5 and placing it alongside Gemini 3.1 Pro and Claude Opus 4.6 on the leaderboard. On Terminal-Bench 2.1, a test of agentic terminal command sequencing, it scored 70.8. On SWE-bench Multilingual, which extends the code evaluation to non-English language repositories, it scored 77.3. These results are Meituan's own reporting; the model's weeks of live exposure on OpenRouter as "Owl Alpha" provides an independent signal that the coding performance claims were observable in real usage before the official benchmarks were published.
Trained on Chinese Chips: The Hardware Significance
LongCat-2.0's most strategically significant claim is that it was trained end-to-end on a cluster of more than 50,000 domestic Chinese ASIC chips. Meituan has not named the specific chip vendor explicitly, but acknowledgement of Huawei's Collective Communication Library in the training infrastructure strongly suggests Huawei Ascend hardware was involved. This makes LongCat-2.0 the first publicly released model of this scale to have been trained and inferred entirely without the Nvidia H100 or H200 hardware that dominates frontier AI training outside China. The practical implication is that US export controls on advanced semiconductors — designed in part to constrain China's AI progress — have not prevented Meituan from producing a model that benchmarks at the frontier tier.
Pricing and API Access
Standard API pricing is set at $0.75 per million input tokens and $2.95 per million output tokens, with a launch promotional rate of $0.30 input and $1.20 output. Cached context reads are free of charge under the standard pricing structure. The model is accessible through Meituan's OpenAI- and Anthropic-compatible API endpoints, and through agent harnesses including Hermes and Claude Code. Model weights are available under an MIT licence on both GitHub and Hugging Face.
What LongCat-2.0 Means for Indian AI Teams
For Indian teams building with AI — whether for internal tooling, customer-facing products, or agentic coding pipelines — LongCat-2.0 introduces a credible new option in the frontier model tier at pricing that undercuts GPT-5.5 and Claude Sonnet 5 significantly. At the promotional rate of $0.30 per million input tokens, cost-intensive use cases such as long-context document analysis, large-codebase refactoring, and multi-turn agentic workflows become materially cheaper. The 1-million-token context window is specifically useful for Indian teams working with large codebases or long documentation corpora where full-file context matters.
There is a meaningful caveat: LongCat-2.0 is a Chinese company's model, and enterprise procurement contracts, data residency policies, and government procurement rules in India may restrict its use in regulated sectors. Teams building for banking, insurance, or government clients should verify applicable data handling requirements before integrating. For unregulated commercial applications — developer tooling, content generation pipelines, or software engineering assistants — the pricing and performance profile merits serious evaluation alongside US model alternatives.
The Bottom Line
Meituan released LongCat-2.0 on 30 June 2026: a 1.6-trillion-parameter MoE model with a 1-million-token context window, trained entirely on domestic Chinese ASIC chips and open-sourced under an MIT licence. It scored 59.5 on SWE-bench Pro, edging past GPT-5.5's 58.6. The model operated on OpenRouter as "Owl Alpha" before its public reveal, accumulating real-world usage at scale. API pricing at the promotional rate of $0.30/$1.20 per million tokens significantly undercuts comparable US frontier models. For Indian engineering teams evaluating coding and agentic AI tooling, LongCat-2.0 is the first credible cost-competitive alternative to US frontier models at this capability tier — with data governance requirements warranting review before integration into regulated applications.
Frequently Asked Questions
What is Meituan LongCat-2.0 and when was it released?+
LongCat-2.0 is a 1.6-trillion-parameter Mixture-of-Experts large language model released and open-sourced by Meituan — China's dominant food delivery and e-commerce platform — on 30 June 2026. It uses a sparse MoE architecture with approximately 48 billion parameters active per token and ships with a native 1-million-token context window maintained by a custom linear-complexity attention mechanism called LongCat Sparse Attention. Before the official release, the model was deployed on OpenRouter under the alias 'Owl Alpha', where it topped usage charts. Model weights are available under an MIT licence on GitHub and Hugging Face.
What benchmark scores does LongCat-2.0 achieve?+
Meituan reports LongCat-2.0 scored 59.5 on SWE-bench Pro, an industry-standard benchmark measuring a model's ability to resolve real GitHub software engineering issues — edging past the 58.6 reported for OpenAI's GPT-5.5 and placing alongside Gemini 3.1 Pro and Claude Opus 4.6. It also scored 70.8 on Terminal-Bench 2.1, which tests agentic terminal command sequencing, and 77.3 on SWE-bench Multilingual, which extends the code evaluation to non-English language repositories. These are self-reported results from Meituan; the model's prior OpenRouter exposure as 'Owl Alpha' provides independent usage signal for the coding performance claims.
How was LongCat-2.0 trained without Nvidia GPUs?+
Meituan states that LongCat-2.0 was trained end-to-end on a cluster of more than 50,000 domestic Chinese ASIC chips. The company did not explicitly name the chip vendor, but acknowledgement of Huawei's Collective Communication Library (HCCL) in the training infrastructure strongly suggests Huawei Ascend hardware was the primary compute platform. LongCat-2.0 is the first publicly released model of this scale — 1.6 trillion parameters — to claim complete training and inference without Nvidia H100 or H200 hardware, the dominant chips for frontier AI model training outside China.
How is LongCat-2.0 priced and how can developers access it?+
Standard API pricing for LongCat-2.0 is $0.75 per million input tokens and $2.95 per million output tokens, with a launch promotional rate of $0.30 per million input tokens and $1.20 per million output tokens. Cached context reads are free of charge. The model is accessible through Meituan's OpenAI- and Anthropic-compatible API endpoints and through agent harnesses including Hermes and Claude Code. Model weights are available under an MIT licence on GitHub and Hugging Face, allowing self-hosted deployments. The promotional pricing significantly undercuts comparable US frontier models at the same capability tier.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.