On June 30, Chinese tech company Meituan officially unveiled LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts AI model. The model had previously run anonymously on OpenRouter for two months under the alias Owl Alpha, during which it achieved top rankings on several platforms. This is the first trillion-parameter model trained end-to-end on domestic Chinese ASICs, using a cluster of over 50,000 accelerators for pretraining on more than 35 trillion tokens. Meituan claims the training run had no rollbacks or irrecoverable loss spikes, a significant achievement given the instability often seen with unproven hardware. LongCat-2.0 features advanced techniques including sparse attention and N-gram embeddings, and it incorporates specialized systems for tool use, reasoning, and conversation. The model's pricing is highly competitive: $0.75 per million input tokens and $2.95 per million output tokens, far below GPT-5.5 and Claude Sonnet 5 rates. In benchmark tests, LongCat-2.0 scored 59.5 on SWE-bench Pro, ahead of GPT-5.5's 58.6 and Gemini 3.1 Pro's 54.2, though behind Claude Opus 4.7 and 4.8. It also scored 73.2 on FORTE, tied with Claude Opus 4.6 but trailing GPT-5.5's 77.8. The model is available via Meituan's API endpoints compatible with OpenAI and Anthropic, and integrates with agent harnesses like Hermes, Claude Code, and OpenClaw. However, self-hosting options are not yet available, as model weights have not been released.

Key facts

• Meituan unveiled LongCat-2.0, a 1.6T-parameter MoE model, on June 30 after anonymous testing as Owl Alpha.
• First trillion-parameter model trained end-to-end on domestic Chinese ASICs; over 50,000 accelerators used.
• Achieved top rankings: 1st on Hermes Agent, 2nd on Claude Code, 3rd on OpenClaw by call volume.
• Pricing: $0.75/M input, $2.95/M output tokens, undercutting GPT-5.5 and Claude Sonnet 5.
• Scored 59.5 on SWE-bench Pro, ahead of GPT-5.5 (58.6) and Gemini 3.1 Pro (54.2).