LongCat 2.0
#21 in Frontier-Spraakmodellemeituan · v2.0 · siet 2026-06-30 · 16× · tolest 03. Juli 2026
46
Momentum
LongCat-2.0 is a frontier-scale Mixture-of-Experts (MoE) language model by Meituan featuring 1.6 trillion total parameters, with 33–56 billion parameters dynamically activated per token. It natively supports a 1-million-token context window and was trained end-to-end on a 50,000-card domestic Chinese AI ASIC superpod cluster without any Nvidia hardware. Designed specifically for agentic coding tasks, it was open-sourced on June 30, 2026, under the MIT license.
Momentum-Verloop
04.04.03.07.
Features
| Coding Benchmark (SWE-bench Pro %) | 59.5% (SWE-bench Pro) — surpasses GPT-5.5 (58.6%) |
| Context Window (Tokens) | 1,000,000 tokens (1M native) |
| Multimodal (Image/Video) | No — LongCat-2.0 is a pure text/code model. Multimodal capabilities (image, audio, video) are reserved for separate models in the LongCat family (LongCat-Next, LongCat-Flash-Omni). |
| NLP Benchmark Score | Terminal-Bench 2.1: 70.8; SWE-bench Multilingual: 77.3 |
| Price per 1M Input Tokens | Standard: $0.75 / 1M input tokens; Promo (limited-time): $0.30 / 1M input tokens. Cache hits: free. |