Kimi Linear

Kimi Team (Moonshot AI) · 2025-11-01 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE hybrid text-only open-weights large frontier-generalist

Design overview

Kimi Linear is a 48B-parameter sparse MoE (3B active) that interleaves token mixing at a uniform 3:1 ratio: three Kimi Delta Attention (KDA) layers for every one full MLA layer. KDA is a gated-delta-rule linear attention — St = (I − βt kt ktᵀ) Diag(αt) S_{t−1} + βt kt vtᵀ — that fuses Gated DeltaNet's delta-rule recurrence with GLA-style channel-wise forgetting: each of the 128 key/value dimensions has its own decay αt, parameterized by a low-rank projection through a bounded decay function, giving fine-grained control of the fixed 128×128 per-head recurrent state. Queries/keys pass through ShortConv + Swish + L2Norm for eigenvalue stability, and the head output gets RMSNorm plus a low-rank data-dependent Sigmoid gate (which the report ties to alleviating attention sink). The kernel story is the second half of the contribution: by binding the DPLR factors to k (a = b = k), the chunkwise algorithm avoids two secondary-chunking steps and three matrix multiplications, running ~2× faster than general DPLR while remaining mathematically the classical delta rule. The whole model is NoPE — MLA layers carry no positional encoding and KDA's learnable decay acts as a multiplicative positional encoding — which also lets MLA convert to pure MQA at inference. The MoE (8/256 experts plus one shared, first layer dense) and the entire recipe (MuonClip, WSD, K2 corpus, annealing/long-context activation) are inherited from Moonlight and Kimi K2. Released with open kernels, vLLM integration, and checkpoints; the same KDA recurrence later becomes the linear-attention layer of Kimi K3, the atlas's next family record.

What the report claims is novel

Atlas assessment: KDA is a synthesis inside the delta-rule family rather than a new mixing family: Gated DeltaNet's recurrence with GLA-style channel-wise gates and DeltaNet's Householder/delta update, plus a genuinely useful kernel contribution (the a=b=k-constrained DPLR chunkwise algorithm, ~2× over general DPLR). As a system, Kimi Linear follows the inter-layer hybrid paradigm already present in the atlas via ai21-jamba (Transformer+Mamba 52B), but it is the first delta-rule-based (KDA) hybrid in the atlas and the first entry whose linear component beats the full-attention baseline under matched training. The NoPE-global-attention choice is shared with recent hybrids (Falcon-H, SwanGPT, and the report's own ablation vs RoPE) and is well motivated. The 'outperforms full attention' claim is internally consistent (identical recipe, 1.4T matched runs) but validated only at 48B/3B scale. The KDA mechanism is the direct ancestor of the KDA layers in moonshot-kimi-k3, whose report refines the same recurrence; the K3 report in this atlas does not credit Kimi Linear by name, but the mechanism is unambiguous.

Relation to the atlas

Lineage in

Influence out

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyKimi Linear[1]
OrganisationKimi Team (Moonshot AI)[2]
Release date2025-11-01[3]
Report URLhttps://arxiv.org/abs/2510.26692[4]
Licensenot disclosed
Open weightsyes[5]

Scale

Total parameters48B[6]
Active parameters / token3B[7]
Layersnot disclosed
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query headsnot disclosed
KV headsn/a[8]
Head dimension128[9]
Vocabulary sizenot disclosed
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention varianthybrid[10]
Attention layer patternuniform 3:1 interleaving — 3 KDA (linear-attention) layers per 1 full MLA layer, repeated through the stack (N = 3, Figure 3); the first layer is dense (no MoE); 3:1 chosen by hybrid-ratio ablation[11]
Depth mixingsequential-residual
ActivationSwish[12]
Expert count256[13]
Experts per token8[14]
Shared experts1[15]
Routingnot disclosed
Load balancingnot disclosed
Expert granularitynot disclosed
MethodNoPE[19]
RoPE base frequencyn/a[20]
Partial RoPEn/a[21]
TypeRMSNorm[22]
Placementnot disclosed
QK-normno
Attention sinksno
Softcappingno

Context

Trained context length4,096[23]
Deployed context length1M[18]
Extension methodsame annealing schedule and long-context activation phase established in Kimi K2; no RoPE/YaRN adjustments required because the model is NoPE[24]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens5.7T[25]
Data composition1.4T tokens sampled from the Kimi K2 pretraining corpus; the 5.7T run follows the same procedure on the K2 corpus[26]
Curriculum / stagingsame annealing schedule and long-context activation phase established in Kimi K2[26]
OptimizerMuonClip (inherited from Kimi K2)[26]
LR scheduleWSD learning-rate schedule; learning rate 1.1 × 10⁻³[26]
Batch scheduleglobal batch size fixed at 32 million tokens[26]
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[27]
Preference optimisationother[28]
Reasoning trainingRLVR on mathematics, code and STEM prompt sets pre-selected at moderate difficulty; truncated importance sampling to counter train/inference precision mismatch; dynamic KL penalty and mini-batch adjustment to avoid entropy collapse; PTX loss for general-capability retention[29]
Distillationnone disclosed

Modality

Typetext-only
Attachmentn/a[30]

Inference efficiency

KV-cache designhybrid cache: KDA layers keep a fixed-size recurrent state (dk × dv = 128 × 128 per head, independent of sequence length) instead of a growing KV cache; MLA layers (1 in 4) keep the low-rank latent KV cache and can be converted to pure MQA at inference under NoPE — up to 75% KV-cache reduction vs full MLA, up to 6× decoding throughput at 1M context
Quantisation shippednone disclosed
Speculative decodingnone disclosed
Serving optimisationsopen-source KDA kernels integrated with vLLM, drop-in compatible with existing caching and scheduling interfaces; prefill uses the FLOP-intensive chunkwise kernel, autoregressive generation switches to the recurrent kernel; hybrid I/O-bounded decoding approaches a 3:1 efficiency ratio vs full attention[31]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MMLU-Pro (4k)51.0Figure 1, Table 3
RULER (128k)84.3Figure 1, Table 5
HellaSwag (Base)82.9Table 3
ARC-Challenge (Base)67.3Table 3
Winogrande (Base)78.6Table 3
BBH (Base)72.9Table 3
MMLU (Base)73.8Table 3
TriviaQA (Base)71.7Table 3
GSM8K (Base)83.9Table 3
MATH (Base)54.7Table 3
EvalPlus (Base)60.2Table 3
CRUXEval-I-cot (Base)56.6Table 3
CRUXEval-O-cot (Base)62.0Table 3
CEval (Base)79.5Table 3
CMMLU (Base)80.8Table 3
BBH (Instruct)69.4Table 4
MMLU (Instruct)77.0Table 4
MMLU-Pro (Instruct)67.4Table 4
MMLU-Redux (Instruct)80.3Table 4
GPQA-Diamond (Avg@8, Instruct)62.1Table 4
LiveBench (Pass@1, Instruct)45.2Table 4
AIME 2025 (Avg@64, Instruct)21.3Table 4
MATH500 (Acc., Instruct)81.2Table 4
HMMT 2025 (Avg@32, Instruct)12.5Table 4
PolyMath-en (Avg@4, Instruct)43.6Table 4
LiveCodeBench v6 (Pass@1, Instruct)26.0Table 4
MRCR (128k)29.6Table 5
HELMET-ICL (128k)90.0Table 5
LongBench V2 (128k)35.0Table 5
Frames (128k)58.8Table 5
RepoQA (128k)68.5Table 5
Long Code Arena (128k)37.1Table 5
RULER@1M (Instruct, 5.7T)94.8Appendix D, Table 9
MMLU-Pro (Base, 5.7T)54.8Appendix D, Table 8
GPQA-Diamond (Avg@8, Instruct, 5.7T)71.7Appendix D, Table 9

Provenance references

  1. stated · Title
  2. stated · Title byline; Appendix A ('The project is developed at Moonshot AI'); header (github.com/MoonshotAI/Kimi-Linear)
  3. stated · Title page: 'arXiv:2510.26692v2 [cs.CL] 1 Nov 2025'
  4. stated · pipeline
  5. stated · Abstract: 'we open-source the KDA kernel and vLLM implementations, and release the pre-trained and instruction-tuned model checkpoints'; fn 2 (huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct)
  6. stated · Abstract, §5.4 ('48 billion total parameters')
  7. stated · Abstract, §5.4 ('3 billion active parameters per forward pass')
  8. n/a · §4 (MLA layers cache a low-rank latent — no KV heads; KDA layers keep a recurrent state — no KV cache; NoPE MLA converts to pure MQA at inference, still no KV heads)
  9. stated · §4 ('dk, dv represent the key and value head dimensions, which are set to 128 for all experiments')
  10. stated · §4 (layerwise hybrid of KDA linear attention and full MLA global attention)
  11. stated · §4, §5.2, Table 1 (hybrid-ratio ablation: 3:1 best; 0:1, 1:1, 7:1, 15:1 worse)
  12. stated · §4 (q, k, v = Swish(ShortConv(·))); MoE FFN activation not described
  13. stated · §5.4 ('Each model activates 8 out of 256 experts')
  14. stated · §5.4 ('activates 8 out of 256 experts')
  15. stated · §5.4 ('including one shared expert')
  16. stated · §4, §5.4.1
  17. n/a · §4 (no rescaling factor applies; no RoPE)
  18. stated · §5.4.1 ('the final checkpoint supports a context length of up to 1 million tokens')
  19. stated · §4 ('we apply NoPE to all full attention (MLA) layers'); §6.1 (KDA's data-dependent decay acts as a learnable multiplicative positional encoding, so no explicit PE anywhere)
  20. n/a · §4 (no RoPE in the model; 'Kimi Linear (RoPE)' is an ablation baseline only, §5.4)
  21. n/a · §4 (no RoPE)
  22. stated · §4 (head-wise RMSNorm before the output projection)
  23. stated · §5.4.1 ('All models are pretrained using a 4,096-token context window')
  24. stated · §5.4.1, §4
  25. stated · §5.4.1 ('an expanded total of 5.7 trillion tokens to match the pretraining tokens of Moonlight')
  26. stated · §5.4.1
  27. stated · §5.4.2 (multi-stage SFT: broad instruction-following data first, then scheduled reasoning-intensive data; dataset extends the Kimi K2 SFT data)
  28. stated · §5.4.2 (same RL algorithm as Kimi K1.5, with truncated importance sampling, dynamically adjusted KL penalty and mini-batch size; RLVR on math/code/STEM prompt sets; no RLHF/PPO/DPO/GRPO named)
  29. stated · §5.4.2
  30. n/a · modality.type (text-only)
  31. stated · Abstract, §3.1, §6.3