Kimi K3

Kimi Team (Moonshot AI) · 2026-07-27 · analysed 2026-08-08 · skill v0.3.5

sparse-MoE hybrid multimodal open-weights frontier frontier-generalist

Design overview

Kimi K3 is a 93-layer, 2.78T-parameter sparse MoE (104.2B active per token) that mixes information along three axes at once. For sequence length, each block interleaves three Kimi Delta Attention layers — a delta-rule linear attention with a channel-wise forget gate, lower-bounded (g_min = −5) decay that keeps chunkwise computation inside the BF16 range, and full-rank output gating — with one Gated MLA layer (DeepSeek-V2-style latent KV, NoPE, full-rank gate) for unrestricted global interaction; 23 blocks plus a final Gated MLA give 69 KDA / 24 MLA layers. For depth, Attention Residuals let each layer attend over prior layer outputs via a learned pseudo-query, partitioned into 8 blocks of 12 layers (plus a partial final block) to bound memory. For width, Stable LatentMoE routes each token through 16 of 896 fine-grained experts operating in a 3584-wide latent space (two full-width shared experts carry common transformations), stabilized by RMSNorm before the up-projection, the bounded SiTU-GLU activation, and Quantile Balancing, an auxiliary-loss-free bias update that sets each expert's bias from the quantile of router-score margins. Native vision arrives through MoonViT-V2, a 27-layer ViT trained from scratch with the next-token objective and projected into the shared embedding space; the 1M context is reached by progressive curriculum (8K→64K pre-training, 256K→1M cooldown) with no positional-encoding changes — the model is fully NoPE, with position carried implicitly by KDA's decay. Post-training runs SFT → domain/effort-specialized RL (general, agentic, coding × low/high/max) → multi-teacher on-policy distillation into one model, with MXFP4 QAT from SFT onward.

What the report claims is novel

Atlas assessment: The block is a deliberate three-axis hybrid (token/depth/channel). vs DeepSeek-V3, the closest atlas peer: same MLA + shared/routed MoE lineage, but K3 replaces the all-softmax-attention backbone with a 3:1 KDA/MLA hybrid, scales routing to 896/16 experts in a latent width (vs 256/8 at full width), and replaces DeepSeek-V3's sign-based bias updates with quantile-based QB. The lower-bounded decay parameterization is a numerical-stability refinement of the delta-rule recurrence family, not a new mixing family. The claimed 2.5× scaling-efficiency gain over Kimi K2 is internal to the report and not independently verifiable. From-scratch vision training with a next-token objective (no contrastive initialization) is the first such claim in the atlas; the report's own evidence is that it matches the SigLIP-initialized baseline with better gradient stability. Depth-wise AttnRes is a genuinely new axis for the atlas (no peer describes attention over layers).

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyKimi K3[1]
OrganisationKimi Team (Moonshot AI)[2]
Release date2026-07-27
Report URLhttps://arxiv.org/abs/2607.24653[3]
Licensenot disclosed
Open weightsyes[4]

Scale

Total parameters2.78T[5]
Active parameters / token104.2B[5]
Layers93[5]
Hidden dimension7168[5]
FFN inner dimension3072[6]
FFN ratio0.4
Query heads96[7]
KV headsn/a[8]
Head dimensionnot disclosed
Vocabulary size160000[9]
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention varianthybrid[10]
Attention layer patternper block: 3 KDA layers then 1 Gated MLA layer, repeated throughout the backbone; 69 KDA + 24 MLA layers total, with an additional Gated MLA at the end of the backbone so the final layer is global attention[11]
Depth mixingattention-residuals[12]
ActivationSiTU-GLU[13]
Expert count896[14]
Experts per token16[15]
Shared experts2[16]
Routingsigmoid router scores with biased top-k selection; mixture weights renormalized over selected experts (bias excluded from weights)[17]
Load balancingQuantile Balancing (QB): auxiliary-loss-free per-expert bias set from the (1 − k/n)-quantile of router-score margins, estimated from per-expert histograms over the global batch; bias frozen at inference[18]
Expert granularityfine-grained LatentMoE: 896 routed experts operate in a compact latent width 3584 (0.5× hidden); 2 full-width shared experts[19]
MethodNoPE[23]
RoPE base frequencyn/a[24]
Partial RoPEn/a[25]
TypeRMSNorm[26]
Placementnot disclosed
QK-normnot disclosed
Attention sinksno
Softcappingno

Context

Trained context length1M[27]
Deployed context length1M[22]
Extension methodprogressive context-extension curriculum (8K → 64K pre-training; 256K → 1M cooldown); no positional-encoding modification required (NoPE)[20]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokensnot disclosed
Data compositioncurated corpus over four text domains — Web Text, Code, Mathematics, Knowledge — plus a large-scale vision corpus (captions, interleaved image–text documents, OCR, perception, video, visual coding); knowledge and mathematics corpora rephrased with style/perspective-diverse prompting and fidelity verification (Kimi K2 recipe); substantial programmatic multimodal data coupling code with rendered visuals (SVG, 3D assets, Webpage, Game, CAD schematics); coordinate supervision in absolute and normalized formats; pipelines build on Kimi K2/K2.5[28]
Curriculum / stagingnative multimodal training from the start (vision and language jointly optimized, interleaved tokens under one next-token objective); progressive context extension (8K → 64K pre-training; 256K → 1M cooldown); long-context data upsampled and synthesized (permuted/concatenated multimodal documents) so tasks require full-context attention[29]
OptimizerPer-Head Muon (Muon with per-head Newton–Schulz orthogonalization for attention projections) with weight clipping (from Kimi K2)[30]
LR schedulecosine decay with 1% linear warmup (scaling-law study favored cosine over Warmup Stable Decay under per-schedule optimal hyperparameters); weight decay 0.1 throughout[31]
Batch schedulenot disclosed
Precisionnot disclosed
ParallelismPipeline Parallelism with virtual stages, Expert Parallelism (MoonEP), ZeRO-1 Data Parallelism, Pipeline ZeRO-2 gradient sharding (CPU offload), and Context Parallelism (KDA Context Parallelism / KCP); all-to-all dispatch overlapped with computation; P2P Muon orthogonalization[32]
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[33]
Preference optimisationother[34]
Reasoning trainingmulti-effort RL: three domain experts (general, general agents, coding agents) × three reasoning-effort levels (low/high/max) with per-problem token-budget control and τ annealing; long-horizon agentic RL at up to 1M context (partial rollouts, persistent sandbox states); verifiable environments: knowledge-graph-guided task synthesis, autonomous execution tasks (AET), kernel optimization, web development, personal assistant workflows, vision-in-the-loop tool use[35]
DistillationMulti-Teacher On-Policy Distillation (MOPD): the nine domain × effort expert policies are consolidated into the unified model via per-token on-policy distillation rewards (clipped log-ratio), with effort-conditioned training[36]

Modality

Typemultimodal[37]
Attachmentnative[38]

Inference efficiency

KV-cache designhybrid cache: KDA keeps a fixed-size recurrent state (d_k × d_v per head) instead of a growing KV cache; Gated MLA keeps the low-rank latent KV cache (per-token latent, reconstructed keys/values); both managed jointly in a unified paged prefix-cache pool with KDA checkpoints at 512-token hash boundaries[39]
Quantisation shippedMXFP4 MoE expert weights with MXFP8 activations at deployment (QAT throughout SFT and RL); all non-expert components (attention projections, latent MoE projections, shared experts, routers) remain in higher precision[40]
Speculative decodingEAGLE-3-style draft model fine-tuned from the pre-trained MTP layer (draft mirrors one backbone block; 7-step unroll; LK loss directly maximizes acceptance rate; draft input fuses features from AttnRes blocks 1, 4, and final)[40]
Serving optimisationsFlashKDA kernels (training/prefill; auto-dispatched flash-linear-attention backend); SM-level intra-device context parallelism; KDA-aware prefix cache with fine-grained (512-token) hash blocks and sparse KDA checkpoints; WarpDecode-based token-centric MoE decoding kernels; cache-aware affinity scheduling with consistent-hashing failover; budget-based admission control; prefill/decode disaggregation[41]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
GPQA Diamond93.5Table 2
CritPt23.4Table 2
AA-LCR74.7Table 2
HLE-Full (w/o tools / w/ tools)43.5 / 56.0Table 2
DeepSWE67.5Table 2
ProgramBench77.8Table 2
Terminal-Bench 2.188.3Table 2
FrontierSWE81.2Table 2
SWE-Marathon42.0Table 2
SciCode58.7Table 2
BrowseComp91.2Table 2
DeepSearchQA (F1)95.0Table 2
ResearchRubrics76.2Table 2
MCPMark-Verified94.5Table 2
MCP-Atlas84.2Table 2
AutomationBench30.8Table 2
SpreadsheetBench 234.8Table 2
OSWorld-Verified84.8Table 2
Harvey Lab-AA (criterion pass rate)94.6Table 2
CorpFin v271.6Table 2
WorldVQA ForceAnswer51.0Table 2
OmniDocBench91.1Table 2
Video-MME (w/ sub)90.0Table 2
MMVU82.1Table 2
MMMU-Pro (w/o tools / w/ Python)81.6 / 83.4Table 2
Math-Vision (w/o tools / w/ Python)94.3 / 97.8Table 2
ZeroBench-main pass@5 (w/o tools / w/ Python)23.0 / 41.0Table 2
Artificial Analysis Intelligence Index v4.157.1Table 5
Vals Index74.7Table 5
WebDev Arena (Elo)1678Table 5

Provenance references

  1. stated · Title
  2. stated · Title page byline; fn 1 (huggingface.co/moonshotai/Kimi-K3)
  3. stated · pipeline
  4. stated · Abstract: 'We release the full Kimi K3 model weights'; fn 1 (HuggingFace)
  5. stated · Table 1
  6. stated · Table 1 (MoE Hidden Dimension per Expert)
  7. stated · Table 1 (Attention Heads)
  8. n/a · §2.1.2
  9. stated · Table 1 (Vocabulary Size 160K)
  10. stated · §2.1 (layerwise hybrid of KDA linear attention and Gated MLA)
  11. stated · §2.1, Table 1 (Attention-Layer Composition)
  12. stated · §2.2 (AttnRes: each layer attends over prior layer outputs with a learned pseudo-query; Block variant partitions 93 layers into 8 blocks of 12 (partial final block), 9 total blocks counting the embedding)
  13. stated · §2.3.2, Table 1
  14. stated · Table 1 (Routed Experts)
  15. stated · Table 1 (Experts Active per Token)
  16. stated · §2.3 (N_s = 2 full-width shared experts in every layer)
  17. stated · §2.3.3 Eq. 13 (s_i = Sigmoid(W_r x_i); argtop_k(s + b))
  18. stated · §2.3.3
  19. stated · §2.3, Table 1 (Latent MoE Dimension 3584)
  20. stated · §3.4
  21. n/a · §3.4 (no rescaling factor applies; no RoPE)
  22. stated · §3.4 (supports a context window of up to 1 million tokens)
  23. stated · §2.1.2, §3.4 (no explicit positional embedding; MLA layers are NoPE; KDA encodes position implicitly through recurrent gating/decay)
  24. n/a · §3.4 (no RoPE anywhere in the model)
  25. n/a · §3.4 (no RoPE)
  26. stated · §2.1.1 (head-wise RMSNorm on KDA output), §2.2 (RMSNorm on AttnRes keys), §2.3.1 (RMSNorm in LatentMoE), §2.4 (MoonViT-V2)
  27. stated · §3.4 (window grows 8K → 64K pre-training, 256K → 1M cooldown), Table 1 (Training Context Length 1M)
  28. stated · §3.1
  29. stated · §3.3, §3.4
  30. stated · §2.5, §3.3
  31. stated · §3.2, §3.3
  32. stated · §5.2
  33. stated · §4.1.1 (three-stage pipeline begins with SFT; expanded agentic SFT dataset synthesized with prior Kimi models plus multi-stage verification and human-in-the-loop annotation; XTML chat template; QAT from the SFT stage onward)
  34. stated · §4.1.2 (policy optimization follows the Kimi K2.5 algorithm with per-token regularization for stale-data tolerance; agentic generative reward model (GRM) with tournament-style group reward and binary comparisons, verbosity-budgeted; report never names RLHF/PPO/DPO/GRPO)
  35. stated · §4.1.2, §4.2
  36. stated · §4.1.3
  37. stated · §2.4 (text, images, and videos processed by a single shared backbone within one context)
  38. stated · §2.4 ('natively multimodal… no post-hoc modality-alignment stage'); MoonViT-V2 (27-layer ViT, ~0.4B/401M params, patch 14, 12 heads) trained from scratch with next-token prediction + lightweight MLP projector; pixel-shuffle 2×2 token downsampling (up to 3584×3584 px); factorized intra-frame spatial / inter-frame temporal attention with temporal pooling; language and vision jointly optimized from the start of training (§3.3)
  39. stated · §5.1, §5.4.1
  40. stated · §4.1.4
  41. stated · §5.1, §5.4