1.6T total / approximately 48B active per token (sparse MoE)
524,288 tokens in the released configuration; officially evaluated at up to 1M tokens
MIT
Open-weight native multimodal Mixture-of-Experts model designed for long-horizon agentic coding, full-project development, creative design, 3D work, visual programming, research, and general tool use. LongCat-2.5 contains 1.6 trillion total parameters and activates approximately 48 billion parameters per token. The released configuration uses a 38-layer, 8,192-dimensional language backbone with Multi-head Latent Attention, LongCat sparse indexing, 768 routed experts, 128 identity-style zero experts, top-12 expert routing, 2,048-dimensional expert feed-forward layers, a 163,840-token vocabulary, and three Multi-Token Prediction layers. Its native configuration specifies 524,288 positions with DeepSeek-style YaRN scaling, while official evaluations exercise contexts as large as one million tokens. Native-resolution image and video features are converted into multi-level discrete tokens and jointly aligned with text in a shared semantic space. The vision stack contains a 32-layer, 1,280-dimensional encoder, while an eight-stage residual visual quantizer uses shared 16,384-entry codebooks; the released video configuration supports up to 128 sampled frames. LongCat-2.5 was trained with additional long-sequence trajectories, long-horizon reinforcement learning, and reproducible tasks reconstructed from failures encountered by real agents. Specialized agentic, reasoning, and interaction teachers were combined using full-vocabulary KL distillation and Tree Attention to improve multi-turn execution, instruction following, reasoning, and safety. Long-context supervised training uses supernode-aware parallelism, optimized sharding, and computation-communication overlap, which Meituan reports delivered 86% greater per-device throughput than its DDP baseline. The training and rollout systems maintain bitwise-identical token probabilities across training, prefill, and decoding, while KV-cache optimizations and LongCat Sparse Attention nearly double tested decode KV capacity without increasing time per output token. Developer-reported scores include 88.4 on Terminal-Bench 2.1, 38.5 on Terminal-Bench 4.0, 64.6 on DeepSWE v1.1, 82.1 on PaperBench, 91.9 on BrowseComp, 63.1 on Humanity's Last Exam with tools, 83.1 on MCP Atlas, 83.3 on LongCatClawBench, 55.4 on 3DCodeBench, 78.6 on Vision2Web, 85.2 on VideoMME, and 72.1 on LongCatClawBench-MM. The chat template supports system prompts, structured tool calls, image inputs, retained reasoning history, and both thinking and non-thinking operation. The full checkpoint comprises 208 principal BF16 shards plus two multimodal shards, totaling approximately 3.562 TB. Meituan also released a W4A4 checkpoint totaling approximately 1.336 TB, quantizing routed-expert weights and activations to four bits for native INT4 execution on domestic AI accelerators. Against its W8A8 baseline, Meituan reports that W4A4 improves prefill throughput by 43.3% and decode throughput by 42.77%, reduces time to first token by 15.91% and time per output token by 19.72%, and produces a 1.73-times grouped-GEMM speedup. Combining W4A4 with an eight-bit KV cache reportedly provides 92% greater decode throughput than W8A8 with a 16-bit KV cache. Both checkpoints are released under the MIT License.