LONGCAT-2.0 — KEY FACTS
● Developer: Meituan LongCat (Chinese tech conglomerate Meituan)
● Parameters: 1.6T total MoE — ~48B active per token (33B-56B range)
● Trained on: 50,000 domestic Chinese ASIC chips — first 1T+ model trained end-to-end on domestic compute
● Context window: 1M tokens native
● License: MIT — unrestricted commercial use, closed-source derivatives permitted
● SWE-bench Pro: 59.5% — beats GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%)
● The reveal: Had been running anonymously as "Owl Alpha" on OpenRouter for 2 months, ranking #1 on Hermes Agent workspace by call volume before identity was disclosed
● Available: longcat.ai, OpenRouter, MIT weights on Hugging Face and GitHub
The Owl Alpha Reveal — What It Means That It Was Secret for Two Months
For two months, a model called "Owl Alpha" ran on OpenRouter without any branding, affiliation disclosure, or marketing. It ranked #1 on the Hermes Agent workspace by monthly call volume, #2 on Claude Code deployments, and #3 across OpenClaw deployments — a position in the top three by actual developer usage across agent frameworks worldwide, with no one knowing who built it. By the time Meituan LongCat stepped forward on June 30, the model had accumulated approximately 10.1 trillion monthly tokens, averaging 559 billion tokens per day, representing a 242% month-over-month growth in volume.
The stealth strategy was deliberate. Meituan used the anonymous period to validate real-world developer adoption before the formal launch — letting the model earn its reputation through use before attaching a name to it. The result: LongCat-2.0 launched with a proven usage record rather than benchmark claims. Developer communities that had been building on Owl Alpha for weeks already had working integrations, agent frameworks, and production workflows before the identity was revealed. This is a fundamentally different go-to-market from how Western AI labs release models — and it worked.
The Benchmarks — Beating GPT-5.5 at Open-Source Prices
| Benchmark |
LongCat-2.0 |
GPT-5.5 |
Claude Sonnet 5 |
Claude Opus 4.8 |
| SWE-bench Pro |
59.5% ✓ (beats GPT-5.5) |
58.6% |
63.2% |
69.2% |
| Terminal-Bench 2.1 |
70.8% |
78.2% |
80.4% |
— |
| FORTE (office workflow) |
73.2 |
77.8 |
— |
~80+ |
| Context window |
1M tokens ✓ |
1.05M tokens |
1M tokens |
200K tokens |
| License |
MIT ✓ open-source |
Proprietary |
Proprietary |
Proprietary |
The Chinese Compute Story — Why Training on Domestic Chips Matters
LongCat-2.0 is the first trillion-parameter model trained and deployed end-to-end on domestic Chinese AI accelerators — not Nvidia GPUs or any US-designed chip. The cluster: 50,000 domestic ASIC cards. Training a 1.6T MoE model at this scale required solving significant engineering challenges around distributed training stability, fault recovery at 10,000+ card scale, and communication optimization specific to the domestic hardware architecture. Meituan spent three years scaling from thousands to 50,000 cards, systematically addressing hardware failures, communication anomalies, and numerical fluctuations at scale.
The geopolitical significance: US export controls have targeted Nvidia's highest-end chips for China since 2022. The fact that Meituan trained a frontier-competitive model on domestic chips — with SWE-bench Pro scores beating GPT-5.5 — is the most concrete evidence yet that US chip export restrictions have not prevented China from training competitive frontier models. VentureBeat's analysis noted directly: "By locking down Western closed-source models and driving up API costs, the US government has left a wide operational window for global developers seeking affordable, high-performance alternatives like those found in Chinese open-source models such as LongCat-2.0."
Who Should Use LongCat-2.0
Best use case: High-volume agentic coding at open-source prices. At 59.5% SWE-bench Pro with MIT license and 1M context, it offers near-Claude-Sonnet-5 coding performance with no per-token API costs for teams self-hosting. The quality-per-dollar math is competitive for repository-scale coding agent work.
Data residency consideration: LongCat-2.0 runs on infrastructure associated with a Chinese company. Enterprise teams with data residency requirements for US government contracts, healthcare (HIPAA), or financial services (FINRA) need to evaluate whether self-hosting the MIT weights on their own infrastructure resolves the residency concern — or whether the China affiliation creates a compliance issue regardless of hosting location.
Where it trails: General agent benchmarks (FORTE 73.2 vs GPT-5.5's 77.8), Terminal-Bench (70.8 vs Sonnet 5's 80.4), and broad reasoning tasks beyond coding. LongCat-2.0 is a coding-specialist MoE, not a general-purpose frontier model. Route hard reasoning, document analysis, and multimodal tasks to Claude Opus 4.8 or GPT-5.5.
Sources: Meituan LongCat official release · VentureBeat · Yahoo Tech · Design For Online · Related: Grok 4 vs Claude Sonnet 5 vs GPT-5.5 benchmarks → · Best AI tools July 2026 rankings → · How to build profitable AI agents →