Meituan Releases LongCat-2.5: A 1.6-Trillion-Parameter Native Multimodal Model Strengthening Long-Horizon Agent Capabilities

Smart Things (Zhidongxi) reported on November 2 that today, Meituan's LongCat team released LongCat-2.5, a native multimodal model for long-horizon agent tasks.

The model has a total parameter count of 1.6 trillion, with about 48 billion parameters activated per token. Building on LongCat-2.0, it focuses on strengthening native visual understanding and agentic coding capabilities.

In several benchmark evaluations published by Meituan, including the automated task benchmark AutomationBench and the general Agent task benchmark LongCatClawBench, LongCat-2.5 leads models such as Claude-Opus-5 and GPT-5.6-Sol; on the terminal operation benchmark Terminal Bench v4.0, LongCat-2.5 scores 38.5, surpassing GPT-5.6-Sol but still trailing Claude-Opus-5.

In LongCat-2.5's long-sequence training, more long-horizon trajectory data was introduced, and through large-scale long-horizon reinforcement learning, the model's end-to-end engineering capabilities in tasks such as complex codebase development and complete project construction were improved. At the same time, LongCat-2.5 converts visual signals into multi-level discrete tokens, allowing visual and text inputs to be processed uniformly by the same backbone model, further supporting Visual Coding, creative design, and 3D content generation.

Facing the memory limitation of domestic accelerators with less than 80GB per card, LongCat-2.5 rearranges FSDP, EP, and CP parallelism around supernodes, stacking hybrid vocabulary parallelism, MoonEP dynamic expert replication, and dual-shard FSDP, and moves visual encoding out of the training computation graph. LongCat claims that after the above optimizations, this FSDP configuration improves single-card throughput in long-sequence SFT tasks by 86% compared with the DDP baseline.

After the model was released, Zhidongxi also immediately conducted hands-on tests around visual coding and office tasks. From the hands-on experience, the model generates quickly and performs well in coding and visual understanding: front-end visual effects are good, and PPT charts are rich, but there is still room for improvement in image details and spatial and layer relationships, and PPT layout can also be further optimized.

First, we asked the model to create an HTML animation of a "pelican riding a bicycle." The animation generated by LongCat-2.5 is highly complete, with a prominent subject, consistent visual style, and relatively rich details. However, the pelican's scarf is slightly crooked, and there is some room for improvement in image details.

After that, we asked LongCat-2.5 to interpret Amazon's third-quarter financial report and generate a PPT with charts.

The analysis framework built by LongCat-2.5 is relatively complete, and the combination of trend charts, structure diagrams, and risk disclosures is also relatively natural. The model also provides investment highlights and risk disclosures based on the financial report. After a preliminary comparison, we did not find data errors. However, some pages of the PPT generated by LongCat-2.5 have overly high information density, some text font sizes are too small, and individual conclusions are somewhat absolute; the layout can still be further optimized and adjusted.

1. Natural and interactive animation, task scenarios cover software development and professional office work

The official LongCat team also provided several cases, highlighting the model's capabilities in software development and professional office work.

In software development, the model can generate real-time fluid simulation and visual effects generation tools, building interactive web pages containing high-frame-rate fluid animations. After clicking or dragging on the interface with the mouse, an ink-flow animation can be formed.

LongCat-2.5 can also develop 3D cultural relic display pages. Combining WebGL rendering and front-end interaction technologies, the model generated a 3D model of the Sanxingdui bronze sacred tree. Users can view the cultural relic from different perspectives by rotating, zooming, and double-clicking to focus, and open a cultural relic information panel.

In creative design and 3D scene generation, the model can build a city sandbox, a miniature carousel scene, and a web page themed on traditional Chinese colors, and complete the full process from visual style extraction and page layout to interaction implementation.

In office scenarios, LongCat-2.5 can integrate multi-channel business data, clarify data definitions, and calculate true profit, then further analyze business performance and input-output based on business goals, and generate recommendations for marketing, supply, inventory, and budget allocation.

The model can also analyze customer feedback, identify recurring issues and follow-up update items, rank them by issue severity, urgency, and business goals, and generate an interactive HTML report that supports clicking to view the original feedback. In addition, users can upload venue photos and enter natural language requirements to generate an event invitation web page that includes registration, list viewing, and table export functions.

2. 1.6 trillion parameters, vision and text unified as discrete tokens

LongCat-2.5 adopts a native multimodal architecture, encoding images of different resolutions into discrete token sequences, which are then input into the backbone model together with text tokens.

In the visual encoding stage, LongCat-2.5 follows the method of LongCat-Next, using a semantic alignment encoder (SAE) to generate native-resolution features, and then compressing visual information into multi-level vocabulary tokens through residual vector quantization (RVQ) to alleviate information loss during visual discretization.

The model also introduces an offline visual token preprocessing mechanism. A fixed tokenizer can generate and cache visual TokenIDs in advance, and the backbone model directly reads the preprocessed sequences, simplifying long-sequence packing and splitting and improving load balancing across different tasks during distributed training.

The LongCat team states that the visual codebook embedding and the backbone model are trained jointly. On a relatively small-scale MoE model with an average of about 3 billion activated parameters, visual and text embeddings show an interwoven distribution in t-SNE visualizations, indicating that the two types of signals have begun to fuse within a shared semantic space.

3. GitHub PRs automatically generate interactive environments, and long-horizon Agentic RL improves end-to-end engineering

LongCat-2.5 has built a complete Agentic RL training pipeline around long-horizon Agent tasks, covering data, environments, algorithms, and the Harness, among other aspects.

At the data level, the LongCat team obtains code tasks from public GitHub repositories and internal real-world scenarios; for filtered PRs and associated issues, it automatically generates interactive environments and verifiable test cases, and verifies whether tasks are valid through fail-to-pass and pass-to-pass tests.

During RL training, the model's real-time rollout trajectories and task pass rates are used for secondary data annotation and filtering. The team also continuously revises and expands F2P test cases based on false-negative and false-positive analysis, improving the coverage of tests over task functionality descriptions and the accuracy of verification signals.

To address the problem of some trajectories taking too long in long-horizon tasks, LongCat-2.5 adopts a partial rollout asynchronous training strategy. Each RL step maintains a certain oversampling rate; once a batch is complete, unfinished trajectories are paused and resumed in the next round of sampling. The system dynamically adjusts the oversampling rate according to task pass rates, while filtering out zero-gradient sample groups that are all correct or all wrong, so as to ensure effective gradients during training.

Since tokens within the same trajectory may come from different historical checkpoints, training also faces the problem of sample staleness. To this end, LongCat-2.5 clips the Logprob difference between the sampling stage and the current training stage, and adjusts each token's gradient contribution according to the accumulated historical difference, reducing the gradient bias caused by stale samples.

In reward design, the model restricts speculative behavior by disconnecting the network and clearing git history and file caches, and introduces an LLM judge to supervise complete trajectories. For behaviors such as repeated thinking and repeated tool calls, the system imposes additional penalties; for overthinking and excessively rapid growth in thinking length, it suppresses them by redistributing positive-sample Reward.

In addition, the LongCat team adopts a mixed training approach across tasks and Scaffolds, enabling the model to adapt to similar tasks and different Harnesses. Compared with the cold-start stage, which covers only a small number of adaptation cases, this training approach can further improve the model's generalization ability across different task environments.

4. Expert group mechanism covers Agent, reasoning, and interaction capabilities

To integrate capabilities from different domains into the mainline model, LongCat-2.5 continues the expert group mechanism of LongCat-2.0, separately optimizing Agent capability, reasoning capability, and interaction experience, covering scenarios such as multi-turn task execution, single-turn reasoning, fine-grained instruction following, and security defense.

In MOPD training, the student model samples online across tasks in different domains, and domain experts train the student model based on the full-vocabulary KL divergence of the student's sampled results.

To reduce the distribution shift between teacher trajectories and student samples, LongCat-2.5 further splits each round of teacher responses into per-cut prefixes of fixed length, and limits the student model's self-sampling length to the corresponding range. During training, multiple Prefixes of the same teacher trajectory are grouped and sent to the same inference engine, sharing Prefill computation and KV Cache.

The LongCat team also builds a training tree with teacher trajectories as the trunk and student sampling results as branches, and, combined with Tree Attention, completes OPD training for the same group of samples through a single forward and backward computation, thereby reducing the repeated computation and storage overhead caused by teacher trajectories.

5. Parallel strategy redesigned around supernode architecture, per-GPU throughput up 86% over baseline

LongCat-2.5 adopts long-sequence training in both the pretraining and supervised fine-tuning stages. In response to the limitation that domestic AI accelerators have less than 80GB of memory per GPU, the training system redesigns its parallel strategy around the supernode architecture.

LongCat-2.5 configures parallel groups according to the supernode topology, with FSDP and EP parallelism both kept consistent with the number of GPUs in a single supernode, and CP parallelism not exceeding that number, so that parameter AllGather, expert Token AlltoAll, and KV AllGather are completed as much as possible within the high-bandwidth interconnect domain inside the supernode.

The system also uses hybrid vocabulary parallelism to shard the base word embedding layer, LMHead, and N-gram embedding layer, further reducing memory usage. To address load imbalance in large-scale expert parallelism, LongCat-2.5 introduces a threshold-triggered dynamic expert replication mechanism based on MoonEP, enabling replication only when load imbalance reaches a preset threshold, so as to balance communication cost and load benefits.

By generating image tokens in advance, the training computation graph no longer includes the visual encoder. Based on the known number of image tokens, the system can estimate in advance the multimodal load ratio during data packing and shuffling, making the modality proportions on different data parallel ranks more similar.

LongCat-2.5 also introduces dual-shard FSDP. Model parameters are sharded within the FSDP domain inside the supernode, while the FP32 master weights and optimizer states are further sharded across FSDP domains through HSDP replicas, reducing resident memory while maintaining parameter communication efficiency. The system reduces host-side overhead through batched execution of gradient operations, merged metric reductions, CPU core binding, and NUMA-aware scheduling.

LongCat states that, after the above optimizations, this FSDP configuration improves per-GPU throughput by 86% over the DDP baseline in long-sequence SFT tasks.

Conclusion: Native multimodal models are beginning to advance toward long-horizon engineering execution

Judging from the technical roadmap of LongCat-2.5, the focus of model upgrades has shifted from simply scaling up parameters to systems engineering such as unified visual modeling, long-horizon Agent training, training-inference consistency, and adaptation to domestic compute.

Especially in tasks such as complex codebase development, 3D page generation, and multi-turn tool calling, whether a model can continuously execute, verify results, and revise strategies based on failed trajectories will become an important metric for measuring Agent capability. LongCat-2.5's combination of native visual understanding with long-horizon Agent training also means that multimodal models are moving further from "understanding content" toward "completing tasks."