Meituan's LongCat team has introduced LongCat-2.5, a native multimodal model built for long-horizon agentic tasks. With 1.6 trillion total parameters and roughly 48 billion activated per token, it builds on LongCat-2.0 with stronger native visual understanding and agentic coding.

Native multimodality through discrete visual tokens

LongCat-2.5 represents visual signals as multi-level discrete tokens, letting the backbone process visual and textual inputs in a single unified token format. It applies Residual Vector Quantization (RVQ) to native-resolution features from a Semantic-and-Aligned Encoder (SAE) — the same approach as LongCat-Next — to overcome the information bottleneck of visual discretization. A fixed tokenizer lets visual token IDs be precomputed and cached offline, so the backbone processes pre-tokenized sequences, which simplifies long-sequence packing and load balancing in distributed training. Visual codebook embeddings are learned jointly with the backbone.

Scaling agentic reinforcement learning

The model is trained with a large-scale, end-to-end agentic RL pipeline spanning data, environments, algorithms, and harnesses for long-horizon tasks. A synthesis pipeline filters GitHub pull requests and linked issues, then uses agents to annotate tasks and build interactive environments, validated with fail-to-pass (F2P) and pass-to-pass (P2P) tests. Training uses asynchronous partial rollouts, disables network access in task environments, removes git history, and uses an LLM judge to audit trajectories for exploits and curb reward hacking. LongCat-2.5 also converts real-world agent failure trajectories into reproducible, verifiable training tasks to close capability gaps.

Trained and served on domestic AI accelerators

LongCat-2.5 was trained on domestic AI accelerators, prioritizing memory efficiency with less than 80 GB per accelerator. Its parallelism is aligned to the supernode topology, and a dual-shard FSDP configuration achieves 86% higher per-device throughput than a DDP baseline in long-sequence SFT. For inference, a W4A4 solution quantizes both expert weights and activations to 4 bits and runs the matrix multiplications on native INT4 units; in online stress tests, prefill throughput improves 43.3% and decode throughput 42.77%, while time-to-first-token and per-token latency drop 15.91% and 19.72% respectively versus a W8A8 INT8 baseline, with a 1.73x average GMM kernel speedup.

Benchmarks and availability

Meituan reports LongCat-2.5 scoring 88.4 on Terminal Bench v2.1, up from LongCat-2.0's 70.8, and publishes a metrics dashboard comparing it against GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8, and Claude-Opus-5. Chat and API access are available now at longcat.ai. The team's Hugging Face organization (meituan-longcat) hosts previous LongCat models, including LongCat-2.0 and LongCat-Next; LongCat-2.5 weights have not been released yet.

The emphasis on domestic accelerators and 4-bit inference is notable — it signals a continued push to run frontier-scale training and serving on non-NVIDIA hardware.

Sources

  • 1.6T total parameters / ~48B activated; native multimodality (RVQ + SAE); agentic RL pipeline; W4A4 inference gains; Terminal Bench v2.1 = 88.4; chat/API available: LongCat blog
  • Hugging Face org (meituan-longcat) hosts previous LongCat models (incl. LongCat-2.0, LongCat-Next); LongCat-2.5 weights not yet released: LongCat on Hugging Face