Meituan LongCat released LongCat-2.5 on November 2. Positioned as a native multimodal model for long-horizon agent tasks, it has 1.6 trillion total parameters and activates about 48 billion parameters per token. Building on LongCat-2.0, it strengthens native visual understanding and agentic coding capabilities. Online demo and API access are now open, and GitHub and HuggingFace weights will be released soon.
On the technical route, LongCat-2.5 represents visual signals as multi-level discrete tokens, follows LongCat-Next's semantic alignment encoder, and applies residual vector quantization. Native-resolution discretization and offline visual token precomputation allow the backbone to uniformly process visual and text inputs. Visual codebook embeddings are jointly learned with the backbone, and semantic-space interleaving is validated on a MoE with about 3B activated parameters.
Agent capabilities rely on an Agentic RL pipeline: a data synthesis pipeline automatically generates training tasks, interactive environments, and verifiable test cases, sourced from GitHub and internal scenarios, and validates them with fail-to-pass and pass-to-pass checks. On the training side, an asynchronous partial rollout strategy mitigates long-tail efficiency issues and suppresses sample staleness. It constrains reward hacking through network disconnection, clearing Git history, and LLM judge supervision. Mixed training across tasks and scaffolds promotes generalization.
The infrastructure is optimized for the limitation that domestic compute single-card VRAM is under 80 GB, using supernode parallelism and dual-sharded FSDP. For long-sequence SFT, single-card throughput is 86% higher than the DDP baseline. On the inference side, the W4A4 scheme quantizes both expert weights and activations to 4-bit. Compared with the W8A8 INT8 baseline, Prefill and Decode throughput increase by 43.3% and 42.77% respectively, TTFT and TPOT decrease by 15.91% and 19.72% respectively, and GMM operators are accelerated by an average of 1.73x.
In evaluations, LongCat-2.5 scores 88.4 and 38.5 on Terminal Bench v2.1 and v4.0, 51.8 on AutomationBench, 91.9 on BrowseComp, 25.5 on Program Bench, 83.1 on MCP Atlas, and 55.4 on 3DCodeBench, with Terminal Bench v4.0 exceeding GPT-5.6-Sol's 37.3. On the vision side, it scores 78.6 on Vision2Web and 96.5 on MathVision.
With a native multimodal architecture and a fully self-developed domestic compute stack, it enters the long-horizon agent competition, continuing the trend of large models shifting from a parameter race to task effectiveness.
For details, see the official technical blog: https://longcat.ai/blog/longcat-2.5.
