LongCat-2.0

Spec

Attribute	Value
Developer	Meituan (via LongCat AI team)
Released	2026-06-30 (open-sourced)
Announced	2026-06-30 (open-source release; ran unannounced on OpenRouter as "Owl Alpha" before the reveal)
Context window	1 million tokens (LongCat Sparse Attention / LSA)
Pricing	unknown
License	MIT (commercially permissive)
Availability	Open-sourced 2026-06-30; previously served on OpenRouter under the codename "Owl Alpha"
Architecture	Mixture-of-Experts (MoE)
Total parameters	1.6 trillion
Active parameters	~33–56B / token
Specialization	Agentic coding
Training hardware	Huawei Ascend 910 (50,000-card cluster) — no NVIDIA chips

Release Context

LongCat-2.0 had been running quietly on OpenRouter under the codename "Owl Alpha" before its identity was revealed at open-source release — it had already reached #1 usage on OpenRouter among agentic coding models before the public announcement.

Technical Highlights

• LongCat Sparse Attention (LSA): a custom linear-complexity attention mechanism enabling the 1M-token context window without quadratic memory growth
• MoE efficiency: 1.6T total parameters but only 33–56B active per token — comparable inference cost to a dense 40–56B model
• Training entirely on domestic Chinese hardware: 50,000 Huawei Ascend 910 chips, in-house parallelism, HCCL library; no U.S.-export-controlled silicon used in training or inference

Significance

1. Chinese hardware independence: First model at 1.6T scale trained and served entirely on non-NVIDIA hardware. Demonstrates that US export controls on AI chips have not prevented frontier-scale training inside China.
2. Open-source Chinese frontier: MIT license makes it freely usable for commercial applications worldwide, including by non-Chinese developers.
3. Agentic coding at 1M context: The context window covers an entire large codebase — relevant for the agentic engineering paradigm where Claude Code and similar tools operate at codebase scale.

Compared To

• GPT-5.6 Sol, Claude Sonnet 5, Gemini 3.5 Pro — all proprietary with access restrictions
• Llama 3.x (Meta), Qwen series (Alibaba), DeepSeek V4 — Chinese/open-source comparable; LongCat-2.0 has the largest parameter count of any publicly open-source MoE model at time of release

Sources

• VentureBeat
• Cryptopolitan
• GeopoliTechs analysis
• sources/blogs/meituan-2026-06-30-longcat-2-0.md

Open Questions

• What are LongCat-2.0's exact benchmark scores vs. GPT-5.5, Claude Sonnet 5?
• Does Meituan plan to release LongCat-2.0 as a service (API)?
• Will the domestic-chip training approach scale to LongCat-3.0?