Models

← All models

LongCat-Flash-Thinking-2601

Meituan · 2026-01-14 · 561.9B parameters

LongCat-Flash-Thinking-2601 is Meituan's open-weight MoE reasoning model, officially reported as 560B total parameters with about 27B activated per token. It extends LongCat-Flash-Thinking with agentic mid-training and large-scale multi-environment reinforcement learning for tool use, search, long-horizon interaction, and robustness to noisy environments; its chat template can explicitly enable or disable reasoning. The checkpoint configuration uses MLA and YaRN to expose a 983,040-token maximum from an 8,192-token original context, while the technical report describes mid-training through 256K context.

Benchmark scores