Kimi K3 · Reference sheet

K3 at a glance

Every number worth remembering, and where in the report to find it. Built to print.

The block

   ┌──────────────────────────────────────────────────────────┐
   │  1 × Gated MLA        + Stable LatentMoE    global, NoPE │
   │  3 × KDA              + Stable LatentMoE    linear, O(n) │  ◀── × 23 blocks
   └──────────────────────────────────────────────────────────┘
                        + 1 trailing Gated MLA

   93 layers  =  69 KDA  +  24 Gated MLA          3:1 hybrid ratio
   AttnRes:   8 blocks × 12 layers (final block partial), N=8
   Input:     MoonViT-V2 (401M, 27 layers) ─▶ MLP projector ─▶ backbone
The K3 block. Three KDA layers to one Gated MLA, repeated 23 times, with a final global layer making 93 in total.

Scale

Total parameters2.78 T
Activated / token104.2 B
Layers93
Hidden dimension7,168
Attention heads96
Vocabulary160 K
Context window1 M
Dense layers1
MTP layers1

Mixture-of-Experts

Routed experts / layer896
Active per token16
Sparsity ratio56
Shared experts2
Latent MoE width ℓ3,584
  as fraction of d0.5 ×
MoE hidden / expert3,072

Vision — MoonViT-V2

Parameters401 M
Layers27
Attention heads12
Patch size14
Pixel-shuffle downsample2 × 2
Max input resolution3584²

Named innovations

NameJob§
KDALinear-cost token mixing2.1.1
Gated MLAGlobal attention, gated output2.1.2
AttnResRetrieve across depth2.2
LatentMoEExperts at half width2.3
SiTU-GLUBounded activations2.3.2
Quantile Bal.Balance 896 experts2.3.3
MoonViT-V2Vision, from scratch2.4
Per-Head MuonPer-head update scale2.5

Key constants

KDA log-decay floor gmin−5
  → min retention α6.7e−3
SiTU gate cap β₁4
SiTU up cap β₂25
  → output bound100
AttnRes blocks N8
Weight decay0.1
LR warmup1%
LR schedulecosine

Training context curriculum

Pre-training start8 K
Pre-training extended64 K
Cooldown start256 K
Cooldown final1 M

Deployment facts

API pricing (per MTok)

Input, cache hit$0.30
Input, cache miss$3.00
Output$15.00

Serving

Recommended accelerators64+
Quantization (QAT)MXFP4 / MXFP8
Coding cache-hit rate>90%
Default effortmax

Headline benchmarks (max effort)

BenchmarkK3best rival
BrowseComp91.290.4
Terminal-Bench 2.188.388.8
ProgramBench77.877.6
SWE-Marathon42.040.0
FrontierSWE81.286.6
DeepSWE67.573.0
WebDev Arena (Elo)1,6781,634
AA Intelligence v4.157.159.9
Caveats that travel with these numbers

All Claude Fable 5 results include potential fallbacks; all GPT-5.6 Sol results include potential cyberguards. K3 leads on BrowseComp, ProgramBench, SWE-Marathon and WebDev Arena, and trails Fable 5 and GPT-5.6 Sol overall. The 2.5× scaling-efficiency claim is the authors' own fit on held-out validation data, not independently replicated.

Stated limitations

LimitationPractical consequence
Sensitivity to thinking historyTrained with preserved thinking history. Quality can become highly unstable if you don't pass reasoning back, or switch sessions mid-stream.
Excessive proactivenessMay make unexpected decisions on your behalf. Constrain via system prompt or AGENTS.md.
User-experience gapNoticeable gap versus Claude Fable 5 and GPT-5.6 Sol, beyond what benchmarks capture.