Meituan-longcat has released LongCat-2.5, a multimodal Mixture-of-Experts model designed for long-horizon agentic tasks. This model boasts 1.6 trillion total parameters, with approximately 48 billion activated per token, and integrates visual understanding with agentic coding capabilities. It has undergone extensive training, including large-scale reinforcement learning and multi-teacher learning, and features optimizations for efficient long-context processing on domestic AI accelerators. AI
IMPACT Sets new SOTA on long-horizon agentic tasks, potentially advancing complex codebase development and creative design.
RANK_REASON Frontier-lab model release with system card [lever_c_demoted from frontier_release: ic=1 ai=1.0]
Read on Hugging Face Trending Models →
- BrowseComp+
- Claude Code
- Claude Opus 4.5
- DeepSWE v1.1
- LongCat-2.0
- LongCat-2.5
- LongCat-Next
- meituan-longcat/LongCat-2.5
- NL2Repo
- Program Bench
- SWE Atlas
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →