Kimi-K3

2026-07-16

Date the weights (or, if earlier, the technical report) were first made public, YYYY-MM-DD, as dated by the publisher. A Hugging Face repository creation date is used only when no dated publisher page exists, with a note.

other (kimi-k3)

License of the released weights as declared by the publisher (model card metadata or license file), e.g. apache-2.0, mit, or the name of a custom license.

text + image + video

Input modalities of the released checkpoint as the publisher states them, from: text, image, video, audio, joined with ' + ' in that order ("text", "text + image + video").

2.78T parameters

Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.

104.2B parameters

Parameters used per token in the forward pass, as reported by the publisher.

Not disclosed

Active parameters excluding the input embedding and the output head (unembedding). Stated by the publisher, or computed from params.active by subtracting vocab × d_model for each of the embedding and output head that params.active is stated to include (basis inferred, note with the arithmetic). Not disclosed when the scope of params.active is not stated.

2.779T parameters

Calculated: embedding + output head + all experts + other weights, moonshotai/Kimi-K3@f831ab6 safetensors headers; reproduces stated total (other modalities) and active (with output head)

Parameters of the text backbone counted from the released checkpoint's safetensors headers (embedding, output head, all experts and other weights; excluding MTP modules, vision/audio towers and n-gram lookup tables). Shown only when the counts reproduce the publisher's stated total and active parameters within 5% under some scope.

2.349B parameters

Calculated: input embedding + output head, moonshotai/Kimi-K3@f831ab6 safetensors headers

Input embedding plus output head (once if tied), counted from the checkpoint.

2.777T parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, moonshotai/Kimi-K3@f831ab6 safetensors headers; reproduces stated total (other modalities) and active (with output head)

Language backbone weights excluding input embeddings, output head, lookup/Engram tables, MTP modules and vision/audio towers. Equals params.counted_total minus params.counted_embedding. Same validation as params.counted_total.

30.23B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, moonshotai/Kimi-K3@f831ab6 safetensors headers

Largest group of language-backbone weights sharing a full layers.<index> tensor path, including every routed expert. Excludes lookup/Engram tables, embeddings, output head, MTP and other modalities. Hybrid checkpoints may store mixers and FFNs as separate layers. All-gather volume also depends on wrapping, sharding and dtype.

103B parameters

Calculated: non-expert weights + routed experts × 16/896, excluding embedding and output head, moonshotai/Kimi-K3@f831ab6 safetensors headers; reproduces stated total (other modalities) and active (with output head)

Active parameters per token excluding the input embedding and output head, counted from the checkpoint as all non-expert weights plus routed-expert weights × top_k / experts. Same validation as params.counted_total.

Table 1 figures (card rounds to 2.8T / 104B). The 401M MoonViT-V2 vision encoder is listed separately; whether the totals include it is not stated.

What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).

93

Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.

7,168

Residual stream width.

163,840

Embedding table rows (vocab_size in config, may include padding).

No

Input embedding and output head share weights.

77.0753

Calculated: d_model / layers

Width-to-depth ratio, d_model divided by the number of blocks.

KDA, MLA

Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).

(3 KDA : 1 MLA) × 23 + 1 MLA

Inferred from the source; see the explanation below.

Token-mixer order, summarized as a pattern or layer positions. The fact's optional sequence lists each mixer in order; a ratio alone does not establish the order. Auxiliary prediction modules and standalone FFNs are excluded.

24

Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.

2.875

Calculated: attn.linear_layers / attn.global_layers

attn.linear_layers divided by attn.global_layers (3 for '3 Gated DeltaNet : 1 GQA').

69

Layers with a fixed-size recurrent state (linear attention, DeltaNet variants, Mamba).

96

Query heads in the global-attention layers.

18,432

Calculated: heads × head_dim

attn.heads × attn.head_dim, the query projection width of global layers.

192

Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).

128

Value dimension per head in global layers (MLA v_head_dim; stated for MLA even when equal to the QK head dim).

512

Width of the compressed KV latent that MLA caches per token per layer.

1,536

Width of the low-rank query projection in MLA; 0 when queries use a full-rank projection.

64

Width of the shared positional key that MLA caches next to the latent.

NoPE (all MLA layers)

Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.

No

Queries and keys are normalized per head before the dot product.

sigmoid, channel-wise full-rank, on MLA (and KDA) output before o_proj

Gate applied to the attention output before the output projection, with its placement.

No

Query/key/value projections have bias terms.

KDA: 96 heads × 128, short conv 4 + Swish on q/k/v, L2-normed q/k, channel-wise decay with log-decay bounded below at −5 (gmin·sigmoid), head-wise RMSNorm + full-rank sigmoid output gate

Configuration of the linear / SSM layers (e.g. Gated DeltaNet heads and dims, Mamba-2 state size), as one string.

96

Value heads (or SSM heads) in the linear / SSM layers.

128

Head dimension of the linear / SSM layers (value head dim for DeltaNet variants, head_dim for Mamba-2).

4

Kernel size of the causal short convolution in linear / SSM layers or attention projections.

SiTU-GLU (β1=4 gate, β2=25 up)

FFN activation (SwiGLU, GeGLU, ...).

33,792

Intermediate width of the dense FFN layers that are actually executed.

SiTU-GLU soft caps: gate 4·tanh(x/4)·σ(x), up 25·tanh(x/25), |f| ≤ 100; all dense, shared and routed FFNs

Clamping of GLU gate / up activations (e.g. SwiGLU limit 7), with the layers it applies to.

No

FFN projections have bias terms.

4.71429

Calculated: ffn.dim / d_model

ffn.dim divided by shape.d_model.

896

Routed experts per MoE layer.

16

Routed experts selected per token.

2

Always-on experts per MoE layer.

3,072

Intermediate width of one routed expert.

6,144

Inferred from the source; see the explanation below.

Total intermediate width of the shared expert(s).

1

Number of initial blocks that use a dense FFN instead of MoE.

sigmoid

Gating function over expert logits (softmax, sigmoid).

Quantile Balancing: aux-loss-free selection bias; each expert's bias recomputed every step from the (1−k/n)-quantile of its routing margins (global-batch histogram estimate), applied from the next step

Load-balancing method, with coefficients where disclosed.

1

Multiplier applied to the combined routed-expert output (routed_scaling_factor).

Yes

Selected expert weights are renormalized to sum to 1.

none

Expert groups and groups selected per token, or node-limited routing, when used.

3,584

Width of the projected space routed experts operate in (LatentMoE), when smaller than d_model.

2

Calculated: d_model / moe.latent_dim

shape.d_model divided by moe.latent_dim, how much LatentMoE compresses the hidden state before routed experts.

0.428571

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

56

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.037482

Calculated: params.active / params.total

params.active divided by params.total.

1

Additional future tokens predicted during pretraining (0 when the model was trained without MTP).

1

MTP modules with distinct weights. 1 with mtp.depth > 1 means one module is applied recursively.

one layer with the structure of a backbone block

Contents of one MTP module (mixer and FFN type, norms, how the previous hidden state and next-token embedding are combined).

Not disclosed

FFN in the MTP module: 'dense' or 'MoE', with its width when it differs from the backbone.

Not disclosed

MTP modules use the main model's input embedding and output head.

Not disclosed

Weight of the MTP loss, including its schedule.

No

The released checkpoint contains the MTP module weights.

RMSNorm

Normalization layer type in the blocks.

pre

Pre-norm, post-norm, or sandwich (pre and post).

1e-05

Epsilon of the block normalization layers.

Block Attention Residuals: layer outputs summed within 12-layer blocks (8 blocks + embedding = 9 sources); each attention and FFN sublayer input is a softmax-weighted mix of block sums and the current partial sum (learned query, RMSNorm'd keys)

Residual design when it is not a plain sum, e.g. hyper-connections with n streams, gated residuals, depth-scaled residuals, FP32 residual accumulation.

8.192K tokens

Sequence length of the main pretraining stage.

1.049M tokens

Longest sequence length trained on after context extension.

1.049M tokens

Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.

8K → 64K (pre-training) → 256K → 1M (cooldown); no positional-encoding change (NoPE)

How context was extended (stages, RoPE base change, YaRN factor).

Not disclosed

Tokens seen in pretraining including annealing and context extension, as reported.

Muon (per-head Newton–Schulz on Q/K/V projections; K2 QK-Clip weight clipping)

Optimizer used for the bulk of parameters.

Not disclosed

Peak learning rate of the main pretraining stage.

Not disclosed

Learning rate at the end of the decay.

cosine with 1% linear warmup

Learning-rate schedule shape with phase boundaries.

1% of training, linear

Warmup length in steps or tokens.

Not disclosed

Global batch size, including any batch-size ramp: sequences × sequence length (token total), or the token total alone when only that is stated.

0.1

Decoupled weight decay coefficient.

Not disclosed

(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.

Not disclosed

Global gradient-norm clipping threshold.

Not disclosed

Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.

K2 QK-Clip weight clipping; SiTU-GLU and RMSNorm before the latent up-projection against routed-branch activation explosion

Interventions used during pretraining to keep it stable (QK-clip, z-loss, logit caps, spike handling), with coefficients.

native multimodal: text and vision tokens interleaved in one next-token objective from the start; vision encoder trained from scratch

Objectives beyond next-token prediction and MTP (fill-in-the-middle rate, distillation, multimodal tokens from the start).

Web Text, Code, Mathematics, Knowledge plus vision (captions, interleaved image–text, OCR, perception, video, visual coding); K2-style rephrasing of knowledge and math; long documents and videos upsampled and synthetic long-context data in cooldown

Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).

PP · VPP · EP · CP · ZeRO-1 DP

Parallelism layout of the main pretraining run as stated, as degree-annotated terms in the order TP, PP, VPP, EP, CP, DP ("PP16 · EP64 · ZeRO-1 DP"); a term without a stated degree has none ("FSDP").

• pretrain-8k

  Seq len

  8192

  Source: paper · p.12 §3.3–3.4

  Our pre-training begins with a context length of 8k tokens

  Native multimodal (text + vision) from the first step. Token counts per stage not disclosed.

• pretrain-64k

  Seq len

  65536

  Source: paper · p.12 §3.3–3.4

  which is later extended to 64k tokens in a subsequent training phase.

  Native multimodal (text + vision) from the first step. Token counts per stage not disclosed.

• cooldown-256k

  Seq len

  262144

  Source: paper · p.12 §3.3–3.4

  and from 256K to 1M tokens during the cooldown phase.

  Native multimodal (text + vision) from the first step. Token counts per stage not disclosed.

• cooldown-1m

  Seq len

  1048576

  Source: paper · p.12 §3.3–3.4

  and from 256K to 1M tokens during the cooldown phase.

  Native multimodal (text + vision) from the first step. Token counts per stage not disclosed.

Ordered pretraining stages. Each item: name, tokens, seq_len, lr, batch_tokens, batch_sequences, steps, note, quote, and source when it differs from the fact's source.

batch size, LR, TPP and model shape retuned by scaling-law studies; separate search per schedule, cosine chosen over WSD

How peak LR, batch size or other hyperparameters were chosen for the released run (fitted scaling law with its formula, μP transfer from a proxy width, sweep), as stated.

27,648 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 24 × (576 × 16/8) B/entry

Retained KV cache divided by context length, including bounded SWA caches and indexer keys. Exports use uniform BF16 at 131,072 tokens. The website can select uniform payload precision (no quantization overhead) or sourced component formats (including scales and documented padding). This is cache payload, not peak prefill memory or persistent prefix storage.

27,648 bytes

Calculated: 27648 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 24 × (576 × 16/8) B/entry

Long-context global KV bytes per token, including indexer keys; excludes bounded SWA caches. Compression rates and cross-layer cache reuse are included. Exports use uniform BF16; the website's storage selector also applies here. This is the field to compare with global-KV figures in reports, using the matching storage recipe.

265 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 24 × (576 × 16/8) B/entry

kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).

34,778,106 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 24 × (576 × 16/8) B/entry

kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).

3,623,878,656 bytes

Calculated: 27648 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 24 × (576 × 16/8) B/entry

Retained KV cache for one sequence at the selected context, including SWA windows and indexer keys. Exports use uniform BF16 at 131,072 tokens. Excludes recurrent state, compressor tail state, draft/MTP caches, temporary prefill tensors, allocator metadata and persistent prefix snapshots. Does not estimate peak prefill or total serving memory.

44

Artificial Analysis Intelligence Index v4.3.2 of the post-trained model, from a dated Wayback snapshot of its Artificial Analysis page. The note names the variant (reasoning setting).

0.685144

DeepSWE v1.1 pass@1 (Datacurve; 113 tasks, same mini-swe-agent scaffold for every model), as a fraction.