Kimi-K2-Base
Moonshot AI · base checkpoint
Open in the comparison table Hugging Face ↗ Paper ↗
KV calculations use one sequence at 131,072 tokens and 16-bit cache elements. They include bounded sliding-window caches. This is retained cache payload, not peak prefill memory. The comparison table also offers sourced mixed-precision layouts, including quantization scales.
Release
- Released
2025-07-11
Source: blog · kimi.com blog index (Wayback 2026-08-17); Kimi K2 entry
Kimi K2 2025-07-11
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.
- License
other (modified-mit)
Source: model-card · README.md; YAML front matter
license_name: modified-mit
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.
Parameters
- Total params
1.04T parameters
Source: paper · p.6 §2.3, Table 2
Kimi K2 is a 1.04 trillion-parameter Mixture-of-Experts (MoE) transformer model
Abstract and card round to 1T.
Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.
- Active params
32.6B parameters
Parameters used per token in the forward pass, as reported by the publisher.
- Active non-embedding (stated)
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.
- Counted total params
1.026T parameters
Calculated: embedding + output head + all experts + other weights, moonshotai/Kimi-K2-Base@ce72df0 safetensors headers; reproduces stated total (backbone) and active (with embedding and 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.
- Counted embedding + head
2.349B parameters
Calculated: input embedding + output head, moonshotai/Kimi-K2-Base@ce72df0 safetensors headers
Input embedding plus output head (once if tied), counted from the checkpoint.
- Non-emb. params (counted)
1.024T parameters
Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, moonshotai/Kimi-K2-Base@ce72df0 safetensors headers; reproduces stated total (backbone) and active (with embedding and 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.
- Largest layer (counted)
17.06B parameters
Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, moonshotai/Kimi-K2-Base@ce72df0 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.
- Active non-emb. (counted)
30.51B parameters
Calculated: non-expert weights + routed experts × 8/384, excluding embedding and output head, moonshotai/Kimi-K2-Base@ce72df0 safetensors headers; reproduces stated total (backbone) and active (with embedding and 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.
- Counting scope
Table 2 figures; abstract rounds to 1T total / 32B active.
Source: paper · p.1 Abstract
32 billion activated parameters and 1 trillion total parameters
What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).
Shape
- Layers
61
Source: config · /num_hidden_layers
Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.
- d_model
7,168
Source: config · /hidden_size
Residual stream width.
- Vocab
163,840
Source: config · /vocab_size
Embedding table rows (vocab_size in config, may include padding).
- Tied embeddings
No
Source: config · /tie_word_embeddings
Input embedding and output head share weights.
- d_model / layers
117.508
Calculated: d_model / layers
Width-to-depth ratio, d_model divided by the number of blocks.
Attention
- Token mixers
MLA
Source: paper · p.6 §2.3
employing Multi-head Latent Attention (MLA) [44] as the attention mechanism
Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).
- Layer layout
61 MLA
Source: paper · p.6 §2.3, Table 2
#Layers 61 61
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.
- Global-attention layers
61
Source: config · /num_hidden_layers
Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.
- Query heads
64
Source: config · /num_attention_heads
Query heads in the global-attention layers.
- Heads × head dim
12,288
Calculated: heads × head_dim
attn.heads × attn.head_dim, the query projection width of global layers.
- QK head dim
192
Source: config · /qk_nope_head_dim + /qk_rope_head_dim
Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).
- V head dim
128
Source: config · /v_head_dim
Value dimension per head in global layers (MLA v_head_dim; stated for MLA even when equal to the QK head dim).
- MLA KV latent
512
Source: config · /kv_lora_rank
Width of the compressed KV latent that MLA caches per token per layer.
- MLA Q latent
1,536
Source: config · /q_lora_rank
Width of the low-rank query projection in MLA; 0 when queries use a full-rank projection.
- Decoupled RoPE dim
64
Source: config · /qk_rope_head_dim
Width of the shared positional key that MLA caches next to the latent.
- Positional encoding
partial RoPE (MLA decoupled, 64 of 192 QK dims)
Source: code · modeling_deepseek.py; DeepseekV3Attention.forward
q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)
Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.
- RoPE base
50,000
Source: config · /rope_theta
RoPE base frequency of the global layers at release.
- QK norm
No
Source: code · modeling_deepseek.py; DeepseekV3Attention.forward
torch.matmul(query_states, key_states.transpose(2, 3)) * self.softmax_scale
RMSNorm only on the Q and KV latents before up-projection; no per-head Q/K norm. Training bounded logits with QK-Clip weight clipping instead (see train.optimizer).
Queries and keys are normalized per head before the dot product.
- QKV bias
No
Source: config · /attention_bias
Query/key/value projections have bias terms.
Feed-forward
- Activation
SwiGLU
Source: code · modeling_deepseek.py; DeepseekV3MLP.forward
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
act_fn is SiLU (config hidden_act).
FFN activation (SwiGLU, GeGLU, ...).
- Dense FFN dim
18,432
Source: config · /intermediate_size
Intermediate width of the dense FFN layers that are actually executed.
- FFN bias
No
Source: code · modeling_deepseek.py; DeepseekV3MLP.__init__
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
FFN projections have bias terms.
- Dense FFN / d_model
2.57143
Calculated: ffn.dim / d_model
ffn.dim divided by shape.d_model.
Mixture of experts
- Routed experts
384
Source: config · /n_routed_experts
Routed experts per MoE layer.
- Experts per token
8
Source: config · /num_experts_per_tok
Routed experts selected per token.
- Expert dim
2,048
Source: config · /moe_intermediate_size
Intermediate width of one routed expert.
- Leading dense layers
1
Source: paper · p.6 §2.3, Table 2
Number of Dense Layers 3 1
config first_k_dense_replace = 1, moe_layer_freq = 1.
Number of initial blocks that use a dense FFN instead of MoE.
- Router score
sigmoid
Source: config · /scoring_func
Gating function over expert logits (softmax, sigmoid).
- Load balancing
aux-loss-free selection bias (e_score_correction_bias added to sigmoid scores for top-k only)
Source: code · modeling_deepseek.py; MoEGate.forward
scores_for_choice = scores.view(bsz * seq_len, -1) + self.e_score_correction_bias.unsqueeze(0)
config topk_method = noaux_tc; the K2 report does not describe the bias update rule or any auxiliary loss.
Load-balancing method, with coefficients where disclosed.
- Routed scaling factor
2.827
Source: config · /routed_scaling_factor
MoEGate.forward multiplies the renormalized top-k weights by it.
Multiplier applied to the combined routed-expert output (routed_scaling_factor).
- Renormalize top-k
Yes
Source: config · /norm_topk_prob
Applied in MoEGate.forward before the routed scaling factor.
Selected expert weights are renormalized to sum to 1.
- Group-limited routing
none
Source: paper · p.6 §2.3, Table 2
Expert Grouping Yes No
DeepSeek-V3 column says Yes, Kimi K2 says No; config n_group = topk_group = 1.
Expert groups and groups selected per token, or node-limited routing, when used.
- Expert dim / d_model
0.285714
Calculated: moe.expert_dim / d_model
moe.expert_dim divided by shape.d_model.
- Sparsity (experts / top-k)
48
Calculated: experts / top_k
Routed experts divided by experts selected per token.
- Active / total
0.0313462
Calculated: params.active / params.total
params.active divided by params.total.
Multi-token prediction
- MTP depth
1
Source: paper · K3 report p.11 §3.2, Table 1 (Kimi K2 column)
Number of MTP Layers 1 layer 1 layer
The K2 report does not mention MTP.
Additional future tokens predicted during pretraining (0 when the model was trained without MTP).
- MTP modules
1
Source: paper · K3 report p.11 §3.2, Table 1 (Kimi K2 column)
Number of MTP Layers 1 layer 1 layer
MTP modules with distinct weights. 1 with mtp.depth > 1 means one module is applied recursively.
- MTP module
Not disclosed
Contents of one MTP module (mixer and FFN type, norms, how the previous hidden state and next-token embedding are combined).
- MTP FFN
Not disclosed
FFN in the MTP module: 'dense' or 'MoE', with its width when it differs from the backbone.
- MTP loss weight
Not disclosed
Weight of the MTP loss, including its schedule.
- MTP weights released
No
Source: config · /weight_map | has eh_proj
Backbone layers 0–60 only; config /num_nextn_predict_layers 0.
The released checkpoint contains the MTP module weights.
Normalization and stability
- Norm
RMSNorm
Source: code · modeling_deepseek.py; DeepseekV3DecoderLayer.__init__
self.input_layernorm = DeepseekV3RMSNorm(
Normalization layer type in the blocks.
- Norm placement
pre
Source: code · modeling_deepseek.py; DeepseekV3DecoderLayer.forward
hidden_states = self.input_layernorm(hidden_states)
Pre-norm, post-norm, or sandwich (pre and post).
- Norm epsilon
1e-06
Source: config · /rms_norm_eps
Epsilon of the block normalization layers.
Context length
- Pretraining context
4.096K tokens
Source: paper · p.9 §2.5
We pre-trained the model with a 4,096-token context window
Sequence length of the main pretraining stage.
- Trained context
32.77K tokens
Source: paper · p.9 §2.5
followed by an additional 60 billion tokens with a 32k sequence length
Longest length the K2 report states was trained; 128K is reached by YaRN. The K3 report (Table 1) lists K2 training context as 128K.
Longest sequence length trained on after context extension.
- Advertised context
131.1K tokens
Source: paper · p.9 §2.5
To extend the context window to 128k, we employed the YaRN method
config max_position_embeddings = 131072.
Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.
- Extension method
4K → 32K (60B tokens), then YaRN to 128K
Source: paper · p.9 §2.5
followed by an additional 60 billion tokens with a 32k sequence length. To extend the context window to 128k, we employed the YaRN method [55].
Release rope_scaling: type yarn, factor 32 over original_max_position_embeddings 4096, beta_fast = beta_slow = 1, mscale = mscale_all_dim = 1 (softmax scale × (0.1·ln 32 + 1)² in DeepseekV3Attention).
How context was extended (stages, RoPE base change, YaRN factor).
Pretraining recipe
- Pretraining tokens
15.5T tokens
Source: paper · p.9 §2.5
We pre-trained the model with a 4,096-token context window using the MuonClip optimizer (Algorithm 1) and the WSD learning rate schedule [25], processing a total of 15.5T tokens.
15.5T = 10T constant + 5.5T cosine; the 400B anneal and 60B long-context tokens are reported separately.
Tokens seen in pretraining including annealing and context extension, as reported.
- Optimizer
MuonClip (Muon + weight decay + RMS matching + QK-Clip, τ=100)
Source: paper · p.4 §2.1
we train Kimi K2, a large-scale MoE model, using MuonClip with τ = 100
Optimizer used for the bulk of parameters.
- Peak LR
0.0002
Source: paper · p.9 §2.5
The first 10T tokens were trained with a constant learning rate of 2e-4 after a 500-step warm-up, followed by 5.5T tokens with a cosine decay from 2e-4 to 2e-5.
Peak learning rate of the main pretraining stage.
- Final LR
7e-06
Source: paper · p.9 §2.5
the learning rate was decayed from 2e-5 to 7e-6.
Learning rate at the end of the decay.
- LR schedule
WSD: 500-step warmup, constant 2e-4 for 10T, cosine to 2e-5 over 5.5T, then 2e-5 → 7e-6 over the 400B (4K) + 60B (32K) annealing/long-context phase
Source: paper · p.9 §2.5
the WSD learning rate schedule [25], processing a total of 15.5T tokens.
Learning-rate schedule shape with phase boundaries.
- Warmup
500 steps
Source: paper · p.9 §2.5
after a 500-step warm-up
Warmup length in steps or tokens.
- Batch size
67M tokens, constant
Source: paper · p.9 §2.5
the global batch size was held at 67M tokens.
Global batch size, including any batch-size ramp: sequences × sequence length (token total), or the token total alone when only that is stated.
- Batch (tokens)
67M tokens
Source: paper · p.9 §2.5
the global batch size was held at 67M tokens.
Global batch size in tokens during the main pretraining stage, after any ramp (sequences × sequence length when stated in sequences, marked inferred).
- Weight decay
0.1
Source: paper · p.9 §2.5
Weight decay was set to 0.1 throughout
Decoupled weight decay coefficient.
- Adam betas
Not disclosed
(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.
- Grad clip
Not disclosed
Global gradient-norm clipping threshold.
- Training precision
No FP8 compute; MoE up-projection and SwiGLU inputs stored as FP8-E4M3 (1×128 tiles, FP32 scales)
Source: paper · p.8 §2.4.3
Inputs of MoE up-projections and SwiGLU are compressed to FP8-E4M3 in 1× 128 tiles with FP32 scales. Small-scale experiments show no measurable loss increase. Due to potential risks of performance degradation that we observed during preliminary study, we do not apply FP8 in computation.
Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.
- Stability measures
QK-Clip (MuonClip), τ=100, per head; MLA: head-specific q/k content weights × √γ, q rope weights × γ, shared k rope untouched; 12.7% of heads triggered at least once in the first 70k steps, inactive afterwards
Source: paper · p.30 App. D
Initial 70000 steps: 12.7% of attention heads triggered QK-Clip for at least once, clamping Smax to 100.
Per-head weight rescaling is Algorithm 1 (p.4 §2.1); no loss spikes over the run (Figure 3).
Interventions used during pretraining to keep it stable (QK-clip, z-loss, logit caps, spike handling), with coefficients.
- Data mix
15.5T tokens of Web Text, Code, Mathematics, Knowledge; knowledge corpora rephrased by LLM (each at most twice), math rewritten as learning notes and translated to English
Source: paper · p.5–6 §2.2
We extended this method to other large-scale knowledge corpora and observed similarly encouraging results, and each corpora is rephrased at most twice.
Domains: p.6 §2.2 "Pre-training Data Overall"; math rewriting follows SwallowMath.
Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).
- Hardware
NVIDIA H800, 8 GPUs per node, 8×400 Gbps RoCE between nodes
Source: paper · p.7 §2.4.1
Kimi K2 was trained on a cluster equipped with NVIDIA H800 GPUs.
Accelerators used for pretraining, count and duration where stated.
- Accelerators
H800
Source: paper · p.7 §2.4.1
Kimi K2 was trained on a cluster equipped with NVIDIA H800 GPUs.
Count not stated.
Accelerator count and type for the main pretraining run, as the count with thousands separators and the chip's short name ("2,048 H800", "2,048 GH200"). A stated range or bound stays as stated ("2,048–4,096 GH200").
- Parallelism
PP16 · VPP · EP16 · ZeRO-1 DP
Source: paper · p.8 §2.4.2
Pipeline Parallelism (PP) with virtual stages [28, 53, 38, 57, 47, 21], 16-way Expert Parallelism (EP) [39], and ZeRO-1 Data Parallelism [60].
The PP degree ends the previous line, which Figure 7 separates from the quote: 'a combination of 16-way'; VPP degree not stated.
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").
- Stages
main
- Tokens
- 15500000000000
- Seq len
- 4096
- Lr
- 0.0002
Source: paper · p.9 §2.5
The first 10T tokens were trained with a constant learning rate of 2e-4 after a 500-step warm-up, followed by 5.5T tokens with a cosine decay from 2e-4 to 2e-5.
Constant 2e-4 for 10T tokens, then cosine to 2e-5 over 5.5T.
anneal
- Tokens
- 400000000000
- Seq len
- 4096
- Lr
- 2e-05
Source: paper · p.9 §2.5
the model was trained on 400 billion tokens with a 4k sequence length
LR decays from 2e-5 to 7e-6 across the anneal and long-context stages.
long-context
- Tokens
- 60000000000
- Seq len
- 32768
Source: paper · p.9 §2.5
followed by an additional 60 billion tokens with a 32k sequence length
Source: paper · p.9 §2.5
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.
Scaling & HP transfer
- Scaling ladder
small-scale Muon MoE runs: sparsity 8–64 at 8 active + 1 shared expert; heads = layers vs 2× layers at 1.2e20–9.0e20 FLOPs
Source: paper · p.6–7 §2.3, Figures 5–6
Through carefully controlled small-scale experiments, we observe that — under a fixed number of activated parameters (i.e., constant FLOPs) — increasing the total number of experts (i.e., increasing sparsity) consistently lowers both the training and validation loss
Sparsity levels and FLOPs budgets from the Figure 5–6 legends; model sizes are not given.
Smaller models trained to choose the architecture or recipe (sizes, token budgets), as disclosed.
- Tokens / param
14.9038
Calculated: train.tokens / params.total
train.tokens divided by params.total.
- Tokens / active param (non-embedding)
507.985
Calculated: train.tokens / params.counted_active_non_embedding
train.tokens divided by params.counted_active_non_embedding, or by the stated params.active_non_embedding when there is no validated checkpoint count.
KV cache
- KV bytes / token
70,272 bytes
Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (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.
- Global KV bytes / token
70,272 bytes
Calculated: 70272 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (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.
- KV bytes / token / B active
2,156 bytes
Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (576 × 16/8) B/entry
kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).
- KV at context / B active
282,536,552 bytes
Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (576 × 16/8) B/entry
kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).
- KV at context
9,210,691,584 bytes
Calculated: 70272 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (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.
Evaluations
- AA Intelligence Index
13
Source: leaderboard · Artificial Analysis model page /models/kimi-k2, Wayback snapshot 2026-10-19
Kimi K2 scores 13 on the Artificial Analysis Intelligence Index
Score of Kimi-K2-Instruct, post-trained from this base. Variant: Kimi K2. Index v4.3.2.
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).