美团正式发布 LongCat-2.5，这是一个面向长程智能体任务的原生多模态模型。模型总参数量为 1.6 万亿，每个 token 激活约 480 亿参数，在 LongCat-2.0 的基础上进一步强化了原生视觉理解与智能体编程能力。
新模型在长程智能体任务中有明显效果提升，这源于我们在长序列训练中加入了更多长程轨迹数据，并通过大规模长程强化学习，提升模型在复杂代码库开发、完整项目构建等任务中的端到端工程能力。围绕数据合成、环境构建与跨 Harness 混合训练，我们持续扩展这条训练链路。此外，我们也将真实 Agent 的失败轨迹转化为可复现、可验证的训练任务，针对模型能力短板持续补充训练数据。
与此同时，LongCat-2.5 具备原生视觉理解能力，可将视觉信号表示为多层级离散 token，使主干模型以统一的 token 形式处理视觉与文本输入，在 Visual Coding 上表现出色，并支持创意设计和 3D 等任务。
这些升级背后，是持续演进的基础设施：我们优化了国产算力上的长上下文训练与 W4A4 推理，并通过训推数值一致性和长上下文 rollout 优化，支撑大规模长程 RL 的高效运行。
指标看板
为直观呈现 LongCat-2.5 当前的能力水平，我们在下图中选取 GPT-5.5、GPT-5.6-Sol、Claude-Opus-4.8 和 Claude-Opus-5 进行对比。涵盖更多模型的详细评测结果见后文表格。
01 视觉表征建模
LongCat-2.5 将视觉信号表示为多层级离散 token，使主干模型以统一的 token 形式处理视觉与文本输入。这一设计融合了原生分辨率视觉离散化、离线视觉预处理，以及与主干模型联合学习的视觉 embedding，既能支持原生多模态融合，也简化了训练中的视觉计算与负载管理。
离散原生分辨率视觉 Transformer：为使视觉基元建模具备与语言建模类似的 token 级抽象，我们将视觉信号表示为多层级词表 token。沿用 LongCat-Next[1]的方法，对语义对齐编码器（Semantic-and-Aligned Encoder，SAE）产生的原生分辨率特征应用残差向量量化（Residual Vector Quantization，RVQ），解决视觉离散化的信息瓶颈。
离线视觉 token 预处理：为减少在线视觉计算和工作负载异构性，固定的 tokenizer 允许我们离线预计算并缓存视觉 token ID。主干模型因此直接处理预先分词的序列，简化长序列打包与切分，并改善分布式训练的负载均衡。
原生视觉表示：视觉码本 embedding 与主干模型联合学习。在平均激活参数约为 3B 的较小规模 MoE 模型上，初步验证显示，视觉与文本 embedding 在 t-SNE 可视化中呈交织分布，表明两者在共享语义空间中的融合。
02 大规模智能体强化学习
RL scaling 是提升模型智能的关键步骤。我们为 LongCat-2.5 搭建了全面的 Agentic RL 训练链路，从数据、环境、算法和 Harness 等方面对长程任务进行大规模扩展。
大规模数据合成与环境构建：我们构建可扩展的数据合成流水线，自动化生成训练任务、可交互环境和可验证的测试用例，数据来源涵盖 GitHub 公开仓库和内部真实场景的代码任务。对于筛选后的 PR 和关联 issue，利用 Agent 自动打标、合成环境，并进行 fail-to-pass（F2P）和 pass-to-pass（P2P）测试校验。在 RL 训练中，实时 rollout 的轨迹和通过率辅助我们进行二次数据打标和筛选，并基于假阴、假阳分析持续修正和扩充 F2P 测试用例，提高对任务功能描述的覆盖率，和验证信号的准确性。
异步强化学习与稳定性优化：我们采用 partial rollout 异步训练策略缓解长尾效率问题：每个 RL step 保持一定过采样率，凑齐批次后中断、保存未完成轨迹，并在下一采样阶段恢复。过采样率基于任务通过率调整，并过滤全对或全错的零梯度样本组，以保证有效梯度。Partial rollout 中，同一轨迹的 token 可能来自不同历史 checkpoint，带来样本陈旧性。为保证训练稳定性，我们裁剪采样时与当前训练时 logprob 的差异，并基于历史累计差异调整每个 token 的梯度贡献，限制陈旧性带来的梯度偏置；此外，从 infra 视角提供了训推一致性的优化方案。
Reward hacking 与输出行为优化：我们通过在任务环境中断网、清除 git 历史和文件缓存，并引入 LLM judge 监督轨迹，以限制模型的投机行为。在输出行为优化上，我们对重复思考和重复工具调用施加额外惩罚；对过度思考和长度增长过快的问题，则通过重新分配正样本中的 Reward 加以抑制。
混合训练与泛化：模型在 RL 中可自主学习和适配近似任务和 Harness。前置冷启阶段仅需覆盖少量适配 scaffold 用例，在 RL 阶段，通过跨多任务和跨 scaffold 的混合训练可显著提高模型泛化能力，并促进不同任务上的性能提升。
03 多专家在线教学
为将各领域能力融入主线模型，我们沿用 LongCat-2.0 的专家组机制，分别优化 Agent 能力、推理能力和交互体验，涵盖多轮任务执行、单轮推理、细粒度指令遵循，以及不牺牲有用性的安全防御。在 MOPD 中，学生模型在各领域任务中在线采样，领域专家通过采样结果上的全词表 KL 散度[2]训练学生。
样本既要反映学生的实际任务行为，也要让教师给出可信的学习信号。为限制分布偏移，先前工作将教师在 SFT 或 RL 阶段见过的轨迹拆为 per-turn prefix，作为学生采样的上下文。我们进一步将每轮教师回复按固定长度拆为 per-cut prefix，并将学生自采样的最大长度限制为该值，以处理单轮长回复中的分布偏移。
为控制计算与存储成本，同一教师轨迹的 prefix 被分为一组，送至同一推理引擎，共享 prefill 计算与 KV cache。训练时，以教师轨迹为树干、学生采样为分支构建树，结合 Tree Attention[3]，通过单次前传与反传完成同组所有样本的 OPD 训练，避免教师轨迹上的冗余计算与存储。
04 国产算力上的长上下文训练
LongCat-2.5 在中训练和监督微调阶段均采用长序列训练。针对国产算力单卡显存不足 80 GB 的限制，训练系统重点优化显存开销，并围绕超节点架构设计并行策略。
4.1 面向超节点的并行策略
LongCat-2.5 根据超节点拓扑配置并行组，FSDP 和 EP 并行度均与单个超节点的卡数一致，CP 并行度不超过该卡数，使参数 AllGather、专家 token AllToAll 和 KV AllGather 均在超节点内的高带宽互联域中完成。
混合词表并行在 DP 和 CP 联合并行域内对基础词嵌入层、LM head 和 N-gram 嵌入层进行分片，进一步降低显存占用。
针对大规模专家并行中的负载不均衡问题，LongCat-2.5 基于 MoonEP 引入阈值触发的动态专家复制机制。仅当 EP 负载不均衡程度达到预设阈值时才启用复制，以避免额外通信开销超过负载均衡收益，并防止路由 token 引发接收端显存溢出、通信操作引发发送端显存溢出。
提前图像 token 生成使训练计算图不再包含视觉编码器，简化了并行实现。利用已知的图像 token 数量，系统可以在数据打包和打散时预估各模态的负载比例，使各数据并行 rank 上的模态占比接近。
4.2 FSDP 与系统优化
LongCat-2.5 引入双重分片 FSDP，模型参数在超节点内的 FSDP 域中分片，FP32 主权重和优化器状态则通过 HSDP 副本进一步跨 FSDP 域分片，在保持参数通信效率的同时，降低常驻显存。
参数、梯度和优化器状态统一组织为连续的扁平化缓冲区，支持非均匀分片和 Muon 优化器分片。同时，通过直接写入权重梯度、复用临时缓冲区，减少中间拷贝，降低峰值显存占用。
系统采用零拷贝 AllGather/ReduceScatter，消除默认通信算法引入的中间拷贝。EP 将 token 与路由概率打包传输，并采用基于 AIV 的通信实现，缓解 AICPU 任务展开瓶颈。参数预取使 AllGather 与计算重叠执行，同时控制与 CP 的资源竞争；ScMoE 则使 EP 通信与稠密分支计算重叠执行。
系统通过批量执行梯度操作、合并指标归约、CPU 绑核和 NUMA 感知调度，降低主机侧开销。
综合上述优化，该 FSDP 配置在长序列 SFT 任务中的单卡吞吐量较 DDP 基线提升 86%。
05 智能体强化学习基础设施
5.1 训推精度一致性
基于 IsoExec[4]、GLM-5[5] 和 XoRL[6]的工作，我们构建了生产级训推一致性方案，使训练、Prefill 和 Decode 可独立选择并行配置，在性能接近无损的同时，保持逐 token logprob 的 bitwise 一致。训推两端统一采用高效的 batch-invariant 算子并对齐框架计算逻辑，补齐 backward 支持。该方案已在实验和生产任务中默认启用，为算子与框架迭代提供统一的数值回归基准。针对因为不同并行方式导致的训推无法对齐的问题，我们在社区的方案上进一步扩展 TP 和 DCP 的训推一致性支持。
TP：以完整 K 维计算替代分段求和，性能接近无损。训练侧保留权重分片，通过权重 All-Gather 和输入 AlltoAll，独立完成完整 K 维 GEMM，无需对局部结果进行 Reduce-Scatter 求和。Prefill 沿用该方案；Decode 则将 o_proj/fc2 改为 column-parallel，通过输入 All-Gather、完整 K 维 GEMM 和输出 AlltoAll 完成计算，既避免每步聚合完整权重，也避免跨卡数值求和。
DCP：在训练与 Prefill 中复现 Decode 的累加路径。我们在训练 forward 和推理 Prefill 中接入配套的 sparse attention 算子，复现指定 DCP 配置下 Decode 的 KV 分块与归约顺序，对齐了三者的计算精度。
5.2 长上下文 rollout 优化
针对负载可控的长上下文多轮 rollout，我们采用 PD 分离，为 Prefill 和 Decode 分别选择合适的并行策略，并从缓存存储、跨层卸载和请求调度三个方面协同优化：
- 扩大有效缓存容量：Prefill 采用 QCP 按 token 分配计算、KVP 按页分散存储，避免 MLA 下 TP 切分带来的 KV 冗余。HiCache 在 KV 回迁显存后，释放无其他读取引用的主存副本，减少两级缓存重复占用；同时基于 Mooncake 重构权重管理与 Actor offload 方案，为 HiCache 释放更多内存。
- 通过跨层预取隐藏搬运开销：利用 LSA[7] 奇数层复用前一偶数层 top-k 索引的特点，将奇数层完整 KV 卸载至主存，显存仅保留小容量工作缓存。偶数层产生索引后，提前加载下一奇数层所需 KV，并与当前层后续计算重叠，在 TPOT 无损的情况下使 Decode KV 缓存容量接近翻倍。
- 通过容量感知调度避免重算雪崩：Router 根据 L1+L2 缓存容量和请求 token 总量控制请求放行，并优先推进完成度较高的 Group，尽早凑齐训练批次、释放缓存，避免多条长轨迹交替推进时反复淘汰彼此前缀，引发大规模重算。
06 W4A4 高效推理
针对 1.6T 级 MoE 模型，我们结合模型结构与国产 AI 加速器的硬件特性，围绕量化精度、底层算子与推理流水开展协同优化，构建 W4A4 高效推理方案。
该方案将专家权重与激活同时量化至 4 bit，并通过原生 INT4 矩阵乘法单元执行相应的矩阵计算。
在精度方面，以 FlatQuant 的可学习等价变换为基础，进一步优化 MoE 模型的量化精度，在已评测的各项下游任务上保持与量化前模型相当的表现。
在算子与数据路径方面，通过专家共享的输入变换，在 token 复制与分发前完成量化，使同一份 INT4 表示贯穿专家通信与计算。将变换、尺度计算、量化和打包融合到专用算子，并结合权重布局与转置访问优化，减少数据搬运和转换开销。
在系统层面，针对 prefill 与 decode 的执行特点，依据模型的数据依赖和硬件资源约束，优化计算与通信的流水重叠。
在线上压测中，相对 W8A8 INT8 基线，prefill 和 decode 吞吐分别提升 43.3% 和 42.77%，TTFT 和 TPOT 分别降低 15.91% 和 19.72%。在算子层面，GMM 相对 W8A8 实现平均 1.73× 加速。
07 能力演示
08 评测说明
详情参见官网技术博客：longcat.ai/blog/longca…。
引用
- [1] LongCat-Next 方法：arxiv.org/abs/2603.27…
- [2] 全词表 KL 散度：huggingface.co/deepseek-ai…
- [3] Tree Attention：arxiv.org/pdf/1909.07…
- [4] IsoExec：vllm.ai/blog/2026-0…
- [5] GLM-5：github.com/THUDM/slime…
- [6] XoRL：kiddyboots216.github.io/mismatch/
- [7] LSA：arxiv.org/abs/2608.01…
体验入口
- 注册认证即赠千万 token，老用户加赠五百万 token，新模型限时福利已上线！
- 在线体验：longcat.chat
- API 接入：longcat.chat/platform/pr…
- GitHub / Hugging Face：即将上线
| 关注「美团技术团队」微信公众号（meituantech）或访问：tech.meituan.com/，阅读更多技术干货！
| 本文系美团技术团队出品，著作权归属美团。欢迎出于分享和交流等非商业目的转载或使用本文内容，敬请注明“内容转载自美团技术团队”。本文未经许可，不得进行商业性转载或者使用。任何商用行为，请发送邮件至 tech@meituan.com 申请授权。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The Mission's official publish LongCat-2.5, a original multimodular model for a long-range AI agent mission. The total model of parameter count is 1.6 trillion, each token activates approximately 48 billion parameters and further enhances native visual understanding and AI agent programming capabilities based on LongCat-2.0.
The new model has had a significant impact on the long-range AI agent mission, which has resulted from the inclusion of more long-range track data in long-sequence training and the enhancement of end-to-end engineering capabilities of the model in complex code library development, complete project construction and so forth through intensive long-range learning. We continue to expand this training chain around data synthesis, environmental construction and cross-Harness training. In addition, we have converted the real Agent's failure trajectory into a repertoire, verifiable training mission, with continuous updating of the training data for model capability panels.
At the same time, LongCat-2.5 has a primary visual understanding that can be expressed as a multilevel separation token, allowing the backbone model to process visual and text input in a unified token format, to perform on the Vision Coding and to support creative design and 3D tasks.
Behind these upgrades are the infrastructure that continues to evolve: we optimize the long context training and W4A4 reasoning of the country ' s economic capacity, and support the efficient operation of large-scale long-range RLs through training in numerical consistency and long context rollout optimization.
Indicator board
To visualize LongCat-2.5's current level of capability, we compare GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8 and Claude-Opus-5 in the figure below. Detailed findings covering more models can be found in the later table.
01 Visual representation modelling
LongCat-2.5 denotes the visual signal as a multi-level separation token, allowing the backbone model to process visual and text input in a unified token format. This design integrates native resolution visual fragmentation, offline visual pre-treatment and visual embedding, which is learning in conjunction with the backbone model, both to support native multi-modular integration and to simplify visual computing and load management in training.
Separated resolution Visual Transformer: In order to make visual foundational modelling abstract token class similar to linguistic modelling, we express visual signals as multi-level vocabulary. Following the LongCat-Next [1] method, the original resolution signature from Semantic-and-Aligned Encoder, SAE is applied to address the information bottlenecks of visual fragmentation by applying the quantization (Residual Victor Quantization, RVQ).
offline visual token pre-treatment: To reduce online visual computation and work load isomericity, fixed tokenizer allows offline to calculate and cache visual token ID. The backbone model thus directly handles the sequence of prephrases, simplifys long sequence packs and cuts, and improves the load balance of distributed training.
Native visual expression: visual code book embedding learning with backbone models. In the smaller MoE model, with an average activation parameter of about 3B, initial validation shows that visual and text embedding is presented in t-SNE visualization, indicating their integration in shared semantic space.
02 Large-scale AI agent intensive learning
RL scaling is a key step in enhancing model intelligence. We have built a comprehensive Agenic RL training link for LongCat-2.5, which has expanded the long-range mission in terms of data, environment, algorithms and Harness.
Large-scale data synthesis and environmental construction: We build scalable data synthesis water lines, automate training missions, interactive environments and verifiable testing examples, with data sources covering GitHub open warehouse and internal real-life scenarios. For filtered PR and associated issue, use Agent automatically to mark, synthesize, and test and verify fail-to-pass (F2P) and pass-to-pass (P2P). In RL training, real-time rollout tracks and pass rates assist us with secondary data marking and filtering, and the continuous revision and expansion of F2P test cases based on pseudo-vaginal, prostate analysis to increase the coverage of mission function descriptions and verify the accuracy of signals.
Step-by-step learning and stability optimization: We use partial rollout training strategies to mitigate longtail efficiency: each RL step maintains a certain sampling rate, breaks the batch, saves unfinished tracks and recovers at the next sampling phase. The oversampling rate is based on the task pass rate adjustment and filters the sample group of zero gradients, both right and wrong, to ensure a valid gradient. Partial rollout, the token of the same trajectory may come from different historical checkpoints, bringing sample old. In order to ensure the stability of the training, we tailor the differences between the sampling and the current training logprob and adjust the contribution of each gradient token based on cumulative historical differences, limiting the gradient bias associated with old age; moreover, the infra perspective provides an optimisation of training consistency.
Reward hacking and output behavioral optimization: We have limited model speculation by cutting off the network, removing guit history and document cache, and introducing LLM judicial to monitor the trajectory. In terms of output behaviour optimization, we impose additional penalties for repeating thinking and repeating tool call; the issue of overthinking and excessive length growth is suppressed by reallocating Reward from positive Z111Z.
Mixed training and generalization: Models can learn autonomously and fit similar tasks and Harness in RL. The pre-cool induction phase needs to cover only a small amount of scaffold. In the RL phase, a combination of training across multiple missions and scaffold can significantly enhance modelability and enhance performance in different tasks.
03 Multi-expert online teaching
In order to integrate capacities in various fields into the main line model, we have followed the LongCat-2.0 expert group mechanism, which optimizes the Agent capabilities, reasoning capabilities and interactive experiences, covering multiple rounds of mission execution, single-wheel reasoning, adherence to fine particle size directives and security defences that do not sacrifice their usefulness. In MOPD, student models are sampled online in various fields of assignment, and field experts train students through full wordsheet KL dispersion[2] of sampling results.
sample needs to reflect both students ' actual mission behaviour and teachers ' credible learning signals. In order to limit the distribution deviation, previous work had broken down the tracks that teachers had seen at the SFT or RL stages into per-turn prefix, context for student sampling. We have further split the per-cycle of teacher responses to a fixed length to per-cut prefix and have limited the maximum length of student samples to that value to handle distribution deviations in single-cycle responses.
In order to control the calculation and storage costs, the prefix of the same teacher track is divided into groups and sent to the same reasoning engine, sharing prefill calculations with KV Cache. In the course of the training, the branches of a teacher's track for a trunk and a student sample for a tree, combined with a Tree Attention [3], complete the OPD training of all sample members of the same group by means of a single forward and reverse transmission, avoiding redundant calculations and storage on the teacher's track.
04 Long context training in national production
LongCat-2.5 uses long-sequence training at the mid-training and supervisory fine-tuning stages. The training system focuses on optimizing the cost of VRAM and designing parallel strategies around hypernodes, in response to the limitations of 80 GB for VRAM.
4.1 Parallel strategies for hypernodes
LongCat-2.5 In the hypernomination configuration, the FSDP and EP parallels are consistent with the number of cards for a single ultranom, and the CP parallels do not exceed that number, so that the parameters AllGather, expert token AllToAll and KV AllGather are completed in the high bandwidth interface within the hypernom.
The hybrid list of words in parallel with the joint DP and CP parallel field is a fraction of the base word embedding, LM head and N-gram embedding, further reducing the VRAM occupancy.
LongCat-2.5 Dynamic Expert Replicating Mechanism based on MoonEP introduction of thresholds to address load imbalances in parallel between large-scale experts. Reproduction is only enabled when the EP load imbalance reaches the predefined threshold in order to avoid additional communication costs exceeding the load balance gain and to prevent the router token from causing the receiving end VRAM spill and the communication operation from the sending end VRAM spill.
The early image token generation has made the training calculation map no longer containing visual encoders, simplifying parallel realization. Using the number of images that are known token, the system is able to estimate the load ratio of the mosaics when the data are packed and dispersed, bringing the mosaics on the parallel rank.
4.2 FSDP and system optimization
LongCat-2.5 Introduction of dual FSDP, model parameters in the FSDP domain in hypernodes, and FP32 sovereign weight and optimizer status further across the FSDP domain by means of HSDP copy, reducing the permanent presence of VRAM while maintaining the efficient communication of parameters.
Parameters, gradients and optimizers are organized in a continuous flat-down buffer zone, supporting non-equilibrium fractions and Muon optimizer fractions. At the same time, through the direct writing of weight gradients and the reuse of temporary buffer zones, intermediate copies were reduced and peak Q82 Z occupancy reduced.
The system uses zero copies of AllGather/ReduceScatter to eliminate intermediate copies introduced by default communication algorithms. EP Packages token with route probability and uses AIV-based communication to mitigate bottlenecks in the AICPU mission. Parameter pre-empts allGather to overlap with calculation, while controlling resource competition with CP; and ScMoE to overlap between EP communications and dense branch calculations.
The system reduces the cost of the side of the host by batches of gradient operations, consolidation of indicator alignment, CPU binding and NUMA sensoring.
Combining the above-mentioned optimization, the FSDP configuration in the long series SFT task increases single-calorie loads by 86% over the DDP baseline.
05 AI agent Enhanced learning infrastructure
5.1 Consistency of training
Based on IsoExec[4], GLM-5[5] and XoRL[6] work, we have built a production-level training consistency programme that allows training, Prefil and Decode to independently select parallel configurations and maintain the bitwise of token logprob while performance is close and intact. Trains to use an efficient Batch-invariant algorithm to align frame calculation logic to complete backward support. The programme has been introduced by default in experimental and production missions, providing a uniform numerical regression baseline for algorithms and frameworks. In response to problems where training cannot be aligned because of different parallel approaches, we have further expanded TP and DCP training coherence support in community programmes.
TP: A full K-dimensional calculation is used to replace the sum of the fractions and the performance is close to intact. The training side preserves a weight split, which is done independently by weighting All-Gather and entering AlltoAll, and which does not require the partial result to be reconciled. Prefil follows the scheme; Decode replaces o proj/fc2 with column-parallel, and completes the calculation by entering All-Gather, full K V GEMM and output AlltoAll, avoiding the aggregation of full weights at each step and avoiding cross-calculations.
DCP: Recreate Decode 's cumulative path in training and Prefil. We're training forward and préfill to access the accompanying sparse attention algorithms, recapitulating the order of the KV sections and returns assigned to Decode under the DCP configuration, aligning the computational precision of the three.
5.2 Long context rollout optimization
For the long load-controlled context multiwheel rollout, we use the PD separation, select the appropriate parallel strategy for Prefil and Decode, and synergize from the cache storage, interlayer unloading and request movement:
- Expand the effective cache capacity: Prefill uses QCP based on the token distribution, KVP separated by page to avoid the KV redundancy associated with the TP split under MLA. After KV moved to VRAM, HiCache released a primary copy with no other reference to reduce double cache duplicate occupancy; and released additional memory for HiCache based on the Mooncake Reweight Management and Actor offload programme.
- Hiding removal expenses by pre-taking across layers: Using LSA [7] to reuse the characteristics of the previous even layer top-k index to unload the odd layer complete KV to the primary storage, VRAM only keeps small capacity cache. After the even layer produces an index, loads the KV required for the next odd layer in advance and overlaps with the current layer subsequent calculation, nearly doubling the capacity of Decode KV Z96Z without damage at TPOT.
- Avoiding avalanche recalculation through volume sensoring: Router is released in accordance with the L1+L2 cache capacity and request token total volume control request, and priority is given to advancing the more complete Group, aligning training batches as early as possible, releasing cache and avoiding multiple elimination of each other ' s prefixes during multiple long-track rotations.
06 W4A4 Efficient reasoning
For the 1.6T level MoE model, we combine the model structure with the hardware properties of the national AI accelerator, synergizing around quantization accuracy, bottom algorithms and reasoning flow, and constructing the W4A4 efficient reasoning formula.
This program gives expert weights to the activation of both quantization to 4 bits and implements the corresponding matrix calculations through the original INT4 matrix multiplication module.
In terms of precision, based on the learning-equivalent transformation of PlatQuant, the precision of the MoE model is further optimized by quantization, which remains comparable to the pre-quantization model for the various downstream tasks assessed.
In terms of algorithms and data paths, quantization is completed before token copying and distribution through input transformations shared by experts, allowing the same INT4 to indicate cross-expert communications and calculations. Combining transformations, scale calculations, quantization and packings with specialized algorithms and optimizing the weight layout and conversion access to reduce data handling and conversion costs.
At the system level, optimal calculation of overlap with the flow of communications is based on model data dependence and hardware resource constraints for the implementation characteristics of prefill and decode.
In online pressurization, prefill and decode increased by 43.3 per cent and 42.77 per cent, respectively, compared to the W8A8 INT8 baseline, while TTFT and TOT decreased by 15.91 per cent and 19.72 per cent, respectively. At the algorithm level, GMM achieved an average acceleration of 1.73 x against W8A8.
07 Capability Demonstration
08 Statement of assessment
For more details, see the official web technology blog Longcat.ai/blog/longca...
References
- [1] LongCat-Next Method: arxiv.org/abs/2603.27...
- [2] Full vocabulary KL diffusion: hugglingface.co/deepseek-ai...
- [3] Tree Attention：arxiv.org/pdf/1909.07…
- [4] IsoExec：vllm.ai/blog/2026-0…
- [5] GLM-5：github.com/THUDM/slime…
- [6] XoRL：kiddyboots216.github.io/mismatch/
- [7] LSA：arxiv.org/abs/2608.01…
The entrance to experience.
- Registration is a gift token, 5 million token for old users, and the new model limited-timefreebie is online!
- Online experience: longcat.chat
- API access: longcat.chat/platform/pr...
- GitHub / Hugging Face: Coming online
WeChatWeChat Official Account (meituantech) or visit: Tech.meituan.com/, reading more technical dry goods!
| This is from the American Corps Technical Team, whose copyright belongs. Reproduction or use of this work for non-commercial purposes, such as sharing and communication, is welcome. This work shall not be reproduced or used for commercial purposes without a permit. For any commercial act, please send an e-mail to tech@meituan.com for authorization.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.