Meituan's Longcat team open-sourced LongCat-Next, a multimodal model based on MoE architecture that integrates five capabilities: text, visual understanding, image generation, and speech. Its core design, DiNA, achieves unified task processing through discrete tokens, while the dNaViT architecture in the visual domain enables excellent image generation performance. Compared to similar models, LongCat-Next demonstrates leading benchmark performance across various metrics, showcasing its advantages in multimodal understanding and generation.

Cursor released the Composer 2 technical report, introducing a complete training solution for its Kimi K2.5 MoE architecture, including two-stage training and a proprietary benchmark called CursorBench. After training, Composer 2 demonstrated significant performance improvements and outperformed other cutting-edge models in terms of inference costs.

Cursor released the Composer 2 technical report on March 25, revealing the training scheme of the Kimi K2.5 model, which adopts a MoE architecture with parameters reaching 10.4 billion. Training is divided into two stages, using real-world scenario simulation for reinforcement learning. The model ultimately achieved 61.3 points on the CursorBench benchmark, a 70% improvement, with inference costs lower than other large language model APIs.

Meituan's LongCat team open-sourced LongCat-Flash-Prover on March 21st, a 560 billion parameter MoE model focused on Lean4 formal theorem proving. The model comprises three capabilities: automatic formalization, sketch generation, and complete proof generation, combined with reasoning tools and the Lean4 compiler to achieve real-time verification. Training utilizes the Hybrid-Experts Iteration Framework and HisPO algorithm to prevent reward hacking. Benchmark results show that the model sets new records for open-source weight models in automatic formalization and theorem proving.

Mistral AI released the open-source code agent Leanstral, designed specifically for Lean 4 formal verification, capable of generating code and proofs that can be automatically validated. The model employs a sparse MoE architecture, outperforming other leading models, and offers free downloads and API calls.