LongCat-Flash

Meituan LongCat Team · 2026-01-29 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE text-only open-weights large long-context

Design overview

LongCat-Flash-Lite is Meituan's demonstration that parameter budget can be moved out of the MoE layers entirely. The 68.5B-parameter model spends 31.4B of them — 46% of the total — on an N-gram Embedding module: each token's embedding is the average of a base 128K-vocab table plus hashed n-gram sub-tables (polynomial rolling hash, K sub-tables with linear projections back into the model space), so capacity grows with vocabulary rather than compute. Because the extra parameters are looked up in O(1) rather than routed, the model keeps total size while cutting activated parameters, which the authors show pays off most at high sparsity: their scaling study (280M/790M/1.3B activated probes over 300B tokens) finds N-gram Embedding beats parameter-equivalent expert scaling once the total/active ratio exceeds ~20, with the crossover shifting higher in wider models and shrinking in deeper ones. On top of the Longcat-Flash MoE backbone (14 shortcut layers, each with two sub-layers; 256 FFN experts plus 128 zero-experts, top-12), the report adds Embedding Amplification — scaling the embedding output by sqrt(D) before the residual merge so the first attention module does not drown it out. Context reaches 256K via YaRN at the 32K stage and a 128K mid-training phase. Inference leans on the reduced activation: Eagle3 3-step speculative decoding, wide expert parallelism, heavy kernel fusion (including Q-Norm + KV-Norm and a unified router kernel), a specialized N-gram Cache, and PDL overlap. The report defers the attention mechanism itself to the Longcat-Flash technical report, so the sequence-mixing design is not documented here.

What the report claims is novel

Atlas assessment: The genuinely novel axis here is parameter allocation: instead of more experts or a new mixing layer, LongCat-Flash-Lite pushes 46% of its budget into a hashed n-gram embedding table (over 30B parameters), a mechanism with O(1) lookup that never touches compute. The N-gram Embedding mechanism itself is borrowed (Clark et al. 2022 / Huang et al. 2025 Over-Encoding), and the report's contribution is the scaling analysis plus system engineering. vs the atlas: the closest relative is moonshot-kimi-linear-48b-a3b (48B-A3B MoE, long context, agentic-leaning), but that model scales via linear attention while LongCat-Flash-Lite scales via embeddings and dynamic zero-expert activation — a fundamentally different sparsity strategy; the 256K YaRN extension and 128k mid-training confirm the LongCat long-context family angle. The 'superior Pareto frontier' claim rests on in-house 300B-token experiments with intersection points that shift with width — honest but not independently verifiable. Weaknesses: the attention mechanism, hidden dim, heads, optimizer, and training precision are all deferred or undisclosed, so the architecture is only partially documented in this report.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyLongCat-Flash[1]
OrganisationMeituan LongCat Team[2]
Release date2026-01-29
Report URLhttps://arxiv.org/abs/2601.21204[3]
Licensenot disclosed
Open weightsyes[4]

Scale

Total parameters68.5B[5]
Active parameters / token2.9B[5]
Layers28
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query headsnot disclosed
KV headsnot disclosed
Head dimensionnot disclosed
Vocabulary size128000[6]
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention variantnot disclosed
Attention layer patternnot disclosed
Depth mixingsequential-residual
ActivationSwiGLU[7]
Expert count256[8]
Experts per token12[9]
Shared expertsnot disclosed
Routingsoftmax top-k with router scaling; zero-expert selection processed in a single unified kernel (Softmax + TopK + router scaling + zero-expert selection)[10]
Load balancingnot disclosed
Expert granularitynot disclosed
MethodRoPE
RoPE base frequencynot disclosed
Partial RoPEno
TypeRMSNorm[13]
Placementpre[14]
QK-normyes[15]
Attention sinksno
Softcappingno

Context

Trained context length131,072[16]
Deployed context length262,144[12]
Extension methodYaRN implemented during the 32k sequence-length training stage, enabling up to 256k; training curriculum: 8k pre-training (11T) -> 128k mid-training (1.5T) -> SFT[17]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens12.5T
Data compositionsame data recipe as LongCat-Flash-Chat (Meituan, 2025); not detailed in this report[17]
Curriculum / stagingthree stages: pre-training on 11T tokens at 8k sequence length, mid-training on 1.5T tokens with sequence length extended to 128k, then supervised fine-tuning[17]
Optimizernot disclosed
LR schedulenot disclosed
Batch schedulebatch size increased at 420B tokens (Figure 10 caption: 'The loss drop at 420B tokens coincides with the batch size increases')[18]
Precisionnot disclosed
Parallelismwide EP (Expert Parallel) and SBO (Single Batch Overlap) adopted for inference deployment[10]
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[19]
Preference optimisationnone disclosed[20]
Reasoning trainingnone disclosed[21]
Distillationnone (trained from scratch)

Modality

Typetext-only
Attachmentn/a[22]

Inference efficiency

KV-cache designnot disclosed
Quantisation shippedquantized deployment with activation quantization fused into operators; quantization format not named[10]
Speculative decodingEagle3 with a 3-step speculative decoding strategy; draft model uses a conventional embedding layer (no n-gram lookup) and n-gram embeddings are cached during drafting to avoid redundant verification-time computation[23]
Serving optimisationswide EP + SBO; extensive kernel fusion (AllReduce+Residual Add+RMSNorm; AllGather+Q-Norm+KV-Norm; ReduceScatter+RMSNorm+Hidden State Combine; router Softmax+TopK+scaling+zero-expert selection); optimized splitkv-and-combine attention kernel (-50% combine latency); PDL (Programmatic Dependent Launch); N-gram Cache with custom CUDA kernels[23]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
Tau2-Airline (avg@8)58.0Table 2
Tau2-Retail (avg@8)73.1Table 2
Tau2-Telecom (avg@8)72.8Table 2
VitaBench (avg@4)7.0Table 2
SWE-Bench (acc)54.4Table 2
TerminalBench (acc)33.75Table 2
SWE-Bench Multilingual38.1Table 2
PRDBench39.63Table 2
GPQA-Diamond (avg@16)66.78Table 2
MMLU (acc)85.52Table 2
MMLU-Pro (acc)78.29Table 2
CEval (acc)86.55Table 2
CMMLU (acc)82.48Table 2
MATH500 (acc)96.8Table 2
AIME24 (avg@32)72.19Table 2
AIME25 (avg@32)63.23Table 2
BBH (base, @1.3T)43.67Table 1
DROP (base, @1.3T)52.43Table 1
GSM8K (base, @1.3T)50.5Table 1
BigCodeBench (base, @1.3T)36.05Table 1

Provenance references

  1. stated · §6.1 ('LongCat-Flash-Lite adopts the same architecture as Longcat-Flash (Meituan, 2025)'); byline 'Meituan LongCat Team'
  2. stated · Byline
  3. stated · pipeline
  4. stated · Abstract ('we introduce and open-source LongCat-Flash-Lite'); HF link huggingface.co/meituan-longcat/LongCat-Flash-Lite
  5. stated · Abstract, §6.1
  6. stated · §3.2.2 ('n-gram vocabulary size set to 30x the base vocabulary (128k)')
  7. stated · §5.1 Eq. 4 / §5.2 Eq. 5 (FFN output Wd(SiLU(Wg x) ⊙ e) — SiLU-gated FFN in the PLE/PLNE formulation of the model's MLP)
  8. stated · §6.1 ('the MoE module consists of 256 FFN experts and 128 zero-experts')
  9. stated · §6.1 ('each token selects 12 experts')
  10. stated · §6.4
  11. stated · §6.1 ('we implement YARN (Peng et al., 2023) during the 32k sequence length training stage, enabling LongCat-Flash-Lite to handle sequences up to 256k tokens')
  12. stated · §6.1 ('handle sequences up to 256k tokens')
  13. stated · §6.4 (kernel fusion: 'AllReduce + Residual Add + RMSNorm')
  14. stated · §3.3.2 ('For pre-normalization architectures, the contribution of N-gram Embedding through the identity connection (residual branch) inherently diminishes as network depth increases')
  15. stated · §6.4 (kernel fusion: 'AllGather + Q-Norm + KV-Norm')
  16. stated · §6.1 ('mid-training during which the sequence length is extended to 128k'; pre-training itself ran at 8k)
  17. stated · §6.1
  18. stated · Figure 10 caption
  19. stated · §6.1 ('finally trained on SFT data'), §6 ('complete pipeline of pre-training, mid-training, and supervised finetuning')
  20. stated · report describes no RL or preference-optimization stage
  21. stated · no reasoning-specific training described beyond SFT
  22. n/a · modality.type (text-only)
  23. stated · §6.4, §4.2