longcat.aiarticleLongCat-2.5: Native Multimodality for Agentic Tasks
LongCat-2.5 is a native multimodal mixture-of-experts model with 1.6 trillion total and about 48 billion activated parameters, built on LongCat-2.0 for long-horizon agentic coding. The post describes discrete RVQ visual tokens, large-scale asynchronous agentic RL with synthesized F2P/P2P-tested environments, on-policy distillation, train-inference numerical alignment, and W4A4 inference on domestic accelerators. A table compares it with GPT-5.5, Claude-Opus-5, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash.
TL;DR Announces LongCat-2.5, a 1.6T-parameter multimodal MoE agent model, with details on its RL pipeline, discrete visual tokens, infrastructure and benchmarks.
Organisations named here
- DeepSeek
- THUDM
- Meituan
- LongCat
Technologies named here
- LongCat-2.0
- gpt-5.5
- Mooncake
- Muon
- GPT-5.6 Sol
- Claude Opus 4.8
- Claude Opus 5
- GLM-5.3
- Kimi K3
- DeepSeek V4.1 Flash
- LLM
- Opus 5
- Terminal-Bench
- DeepSWE
- AutomationBench
- HLE
- BrowseComp
- CharXiv
- GDPval-AA
- Vision2Web
- 3DCodeBench
- Claw-Eval
- FinSearchComp
- FlatQuant
- FrontierScience
- FSDP
- HiCache
- IsoExec
- JobBench
- LongCat-2.5
- LongCat-Next
- MathVision
- MCP Atlas
- MoonEP
- NL2Repo
- PaperBench
- PosterBench
- Program Bench
- ResearchRubrics
- SkillsBench
- SpreadsheetBench 2
- SWE Atlas
- Tree Attention
- VideoMME
Techniques named here
- Agentic reinforcement learning
- Multi-teacher on-policy distillation
- Prefill-Decode disaggregation
- Asynchronous partial rollouts
- Batch-invariant kernels
- Expert parallelism with dynamic expert replication
- FlatQuant
- Residual Vector Quantization (RVQ)
- Tree Attention
- W4A4 quantization
Concepts named here
- Mixed training across scaffolds
- Native multimodality
- Policy staleness
- Train-inference numerical consistency
Also named here
- Mixture of Experts
- Residual Vector Quantization
Related in models
