K2-Horizon-MoVA-36B-A4B (mid_4_10000)

2026-09-03

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.

apache-2.0

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

Inferred from the source; see the explanation below.

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").

36B parameters

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

4B 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.

Not calculated

Calculated: checkpoint count (37.4B backbone) does not reproduce the stated total and active counts

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.

Not calculated

Calculated: checkpoint count (37.4B backbone) does not reproduce the stated total and active counts

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

Not calculated

Calculated: checkpoint count (37.4B backbone) does not reproduce the stated total and active counts

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.

Not calculated

Calculated: checkpoint count (37.4B backbone) does not reproduce the stated total and active counts

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.

Scope of the 36B and 4B counts is not stated. xLLM's num_parameters formula on this config gives 36.2B total and 4.7B active excluding the embedding and output head (1.28B together).

Inferred from the source; see the explanation below.

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

48

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

2,560

Residual stream width.

250,624

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

No

Input embedding and output head share weights.

53.3333

Calculated: d_model / layers

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

GQA (MoVA: routed value projections in 45 of 48 layers)

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

3 GQA, then 45 GQA with MoVA values

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.

48

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

32

Query heads in the global-attention layers.

8

Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.

4,096

Calculated: heads × head_dim

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

4

Calculated: heads / kv_heads

attn.heads divided by attn.kv_heads in global layers.

128

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

RoPE

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

10,000,000

RoPE base frequency of the global layers at release.

1

Inferred from the source; see the explanation below.

Fraction of each head's QK dimensions that are rotated.

No

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

softplus (β = ln 2) gate per head dimension on the attention output, before o_proj, in all layers

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

No

Query/key/value projections have bias terms.

SwiGLU

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

6,144

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

No

FFN projections have bias terms.

2.4

Calculated: ffn.dim / d_model

ffn.dim divided by shape.d_model.

100

Routed experts per MoE layer.

8

Routed experts selected per token.

1

Always-on experts per MoE layer.

768

Intermediate width of one routed expert.

768

Total intermediate width of the shared expert(s).

3

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

sigmoid

Gating function over expert logits (softmax, sigmoid).

per-expert bias for top-k selection only (update rate 1e-3) plus 'dot' auxiliary loss, coefficient 1e-4

Load-balancing method, with coefficients where disclosed.

2.5

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

Yes

Selected expert weights are renormalized to sum to 1.

0.3

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

12.5

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.111111

Calculated: params.active / params.total

params.active divided by params.total.

0

Inferred from the source; see the explanation below.

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

RMSNorm (2 groups of 1280)

Normalization layer type in the blocks.

pre

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

1e-06

Epsilon of the block normalization layers.

8.192K tokens

Sequence length of the main pretraining stage.

524.3K tokens

Longest sequence length trained on after context extension.

524.3K tokens

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

8K → 32K (1.1T) → 128K (498B) → 512K (110B, then 199B with agentic/reasoning SFT data)

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

24.81T tokens

Inferred from the source; see the explanation below.

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

AdamW

Optimizer used for the bulk of parameters.

0.0004

Peak learning rate of the main pretraining stage.

Not disclosed

Learning rate at the end of the decay.

WSD with no decay phase in pretraining: linear warmup, then constant 4e-4 to step 1,100,000

Learning-rate schedule shape with phase boundaries.

3,000 steps, linear from 1e-4 × peak

Warmup length in steps or tokens.

16 × 8,192 per data-parallel rank; global batch not stated

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

0.06

Decoupled weight decay coefficient.

(0.9, 0.95)

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

1

Global gradient-norm clipping threshold.

BF16

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

Gaussian, std 0.02

Weight initialization (std or scheme) as stated by the publisher; config initializer_range alone is not evidence.

no attention across document boundaries (BOS-delimited segments)

Whether attention crosses document boundaries within a packed sequence.

Same pretraining tokens as K2 Horizon 3.7B, 7B and 32B; Midtraining Stage 4 shifts the mix toward agentic and reasoning SFT data

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

TP2 · FSDP

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").

• pretraining

  Tokens

  22900000000000

  Seq len

  8192

  Steps

  1100000

  Source: model-card · README.md; Training Overview table

  Pretraining | 1100000 | 22.9T | 8K | Pretraining.

• midtraining-1

  Tokens

  1100000000000

  Seq len

  32768

  Steps

  55000

  Source: model-card · README.md; Training Overview table

  Midtraining — Stage 1 | 55000 | 1.1T | 32K | Context extension.

  Context extension.

• midtraining-2

  Tokens

  498000000000

  Seq len

  131072

  Steps

  25000

  Source: model-card · README.md; Training Overview table

  Midtraining — Stage 2 | 25000 | 498B | 128K | Context extension.

  Context extension.

• midtraining-3

  Tokens

  110000000000

  Seq len

  524288

  Steps

  5500

  Source: model-card · README.md; Training Overview table

  Midtraining — Stage 3 | 5500 | 110B | 512K | Context extension.

  Context extension.

• midtraining-4

  Tokens

  199000000000

  Seq len

  524288

  Steps

  10000

  Source: model-card · README.md; Training Overview table

  Midtraining — Stage 4 | 10000 | 199B | 512K | Continued context extension from Stage 3, with the data mix shifted toward agentic and reasoning SFT data.

  Mix shifted toward agentic and reasoning SFT data; final checkpoint is tag mid_4_10000.

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.

689.083

Calculated: train.tokens / params.total

train.tokens divided by params.total.

196,608 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 48 × (2048 × 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.

196,608 bytes

Calculated: 196608 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 48 × (2048 × 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.

49,152 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 48 × (2048 × 16/8) B/entry

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

6,442,450,944 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 48 × (2048 × 16/8) B/entry

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

25,769,803,776 bytes

Calculated: 196608 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 48 × (2048 × 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.

25

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).