Kolibri-1

2026-10-03

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

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").

78.1B parameters

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

3.458B parameters

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

3.13B parameters

Inferred from the source; see the explanation below.

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.

78.1B parameters

Calculated: embedding + output head + all experts + other weights, Aleph-Alpha/Kolibri-1-BF16@cc6d729 safetensors headers; reproduces stated total (backbone) 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.

655.4M parameters

Calculated: input embedding + output head, Aleph-Alpha/Kolibri-1-BF16@cc6d729 safetensors headers

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

77.45B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, Aleph-Alpha/Kolibri-1-BF16@cc6d729 safetensors headers; reproduces stated total (backbone) 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.

1.549B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, Aleph-Alpha/Kolibri-1-BF16@cc6d729 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.

3.13B parameters

Calculated: non-expert weights + routed experts × 6/384, excluding embedding and output head, Aleph-Alpha/Kolibri-1-BF16@cc6d729 safetensors headers; reproduces stated total (backbone) 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.

Total includes the input embedding, the untied LM head and the per-layer expert-bias buffers (equals BF16 index total_size 156,206,149,120 bytes / 2). Active count excludes the input embedding but includes the LM head. No MTP or vision modules.

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

50

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

2,560

Residual stream width.

128,000

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

No

Input embedding and output head share weights.

UniBPE (bottom-up BPE merges chosen by Unigram loss) for 119,744 merges, then 7,900 standard BPE merges, 256 byte-fallback tokens, digits split, 100 reserved/special tokens; trained on a 100 GB slice of the German/English pre-training mix

Tokenizer type and lineage as described by the publisher (e.g. byte-level BPE shared with an earlier model).

51.2

Calculated: d_model / layers

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

GQA, SWA

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

4 SWA : 1 global

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.

10

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

4

Calculated: attn.local_layers / attn.global_layers

attn.local_layers divided by attn.global_layers, the sliding-window to global interleave ratio (5 for '5 SWA : 1 global').

40

Layers with a fixed attention window.

48

Query heads in the global-attention layers.

4

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

6,144

Calculated: heads × head_dim

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

12

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).

513 tokens

Sliding-window size of the local layers.

4

KV heads in sliding-window layers (stated even when equal to attn.kv_heads).

48

Query heads in sliding-window layers (stated even when equal to attn.heads).

128

Query/key head dimension in sliding-window layers. Value head dimensions may differ; see attn.local_v_head_dim.

RoPE (SWA), NoPE (global)

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

10,000

RoPE base frequency of the sliding-window layers (stated even when equal to attn.rope_theta).

Yes

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

No

Query/key/value projections have bias terms.

SwiGLU

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

No

FFN projections have bias terms.

384

Routed experts per MoE layer.

6

Routed experts selected per token.

1

Always-on experts per MoE layer.

512

Intermediate width of one routed expert.

512

Total intermediate width of the shared expert(s).

0

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

sigmoid

Gating function over expert logits (softmax, sigmoid).

Exact Quantile Balancing (expert-bias update to the exact global-batch quantile, recentred, no tunable coefficient) + Load-Error Injection into router-score gradients (λ = 1e-5, soft clamp τ = 1); no auxiliary loss

Load-balancing method, with coefficients where disclosed.

No

Selected expert weights are renormalized to sum to 1.

0.2

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

64

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.0442694

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

Normalization layer type in the blocks.

sandwich

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

1e-06

Epsilon of the block normalization layers.

16.38K tokens

Sequence length of the main pretraining stage.

262.1K 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.

16K (pre-training) → 64K (mid-training) → 256K (long-context stage, 201B tokens); no RoPE base change or scaling, since only SWA layers use RoPE; validated to 1M by extrapolation

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

24T tokens

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

Muon (Nesterov, momentum 0.95, spectral scaling) for attention and expert weights; AdamW for LM head and router; Adam for token embedding and norm scales

Optimizer used for the bulk of parameters.

0.001

Peak learning rate of the main pretraining stage.

0.001

Inferred from the source; see the explanation below.

Learning rate at the end of the decay.

Warmup-Stable-Merge: linear warmup over ~100B tokens (1,325 steps), then constant through pre-training, mid-training and long-context training; no decay, final weights are a checkpoint average

Learning-rate schedule shape with phase boundaries.

1,325 steps (~100B tokens)

Warmup length in steps or tokens.

4,608 × 16,384 (75.5M tokens); mid-training 1,536 × 65,536 (100.7M); long-context 768 × 262,144 (201.3M)

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

75.5M tokens

Inferred from the source; see the explanation below.

Global batch size in tokens during the main pretraining stage, after any ramp (sequences × sequence length when stated in sequences, marked inferred).

0.000244141

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 (bfloat16 parameters, FP32 reductions)

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

Truncated normal (±3σ): Q/K/V and expert up/gate std 1/√d_in, router std 1/d_in, token embedding std 1; attention output, expert down and LM head zero; norm scales 1, post-sublayer norm scales 1/√L; expert biases 0

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

next-token prediction only, on packed 16,384-token sequences

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

20T-token German/English corpus (~62.5% English, ~23.9% German, ~13.6% code): curated Common Crawl, ~24% synthetic (rephrases with Gemma-4-26B-A4B and Mistral-NeMo-12B, translations), STEM, instruction and specialised legal/public-sector data; drawn from a 29.3T-token pool

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

2026-06-18

Pretraining data cutoff date.

768 NVIDIA B200 (96 HGX 8×B200 nodes, InfiniBand); pre-training 21 days (511 h, 392k GPU-hours); mid-training 5 days (90k GPU-hours); long-context 13 h (10k GPU-hours)

Accelerators used for pretraining, count and duration where stated.

768 B200

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").

EP8 · FSDP16 · DP6

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").

6.4e23 FLOPs; 392k GPU-hours pre-training + 90k mid-training + 10k long-context

Pretraining compute as stated (GPU-hours or FLOPs).

uniform average of the last 20 long-context checkpoints, 10B tokens apart

EMA or checkpoint merging applied to produce the released base weights.

• pre-training

  Tokens

  20000000000000

  Seq len

  16384

  Lr

  0.001

  Batch sequences

  4608

  Batch tokens

  75497472

  Steps

  264750

  Source: model-card · README.md; Training Details, Pre-training

  The pretraining phase covered 20T tokens over 264,750 optimization steps.

  Muon LR after 1,325 warmup steps; batch_tokens is 4,608 × 16,384 (card: 75.497 million).

• mid-training

  Tokens

  3440000000000

  Seq len

  65536

  Lr

  0.001

  Batch sequences

  1536

  Batch tokens

  100663296

  Steps

  34200

  Source: model-card · README.md; Training Details, Pre-training

  Its 3.44 trillion token budget corresponds to 34,200 optimization steps.

  Higher-quality reasoning, math, code and instruction data; resumes the optimizer state with no warmup, same constant LR. batch_tokens is 1,536 × 65,536 (card: 100.663 million).

• long-context

  Tokens

  201000000000

  Seq len

  262144

  Lr

  0.001

  Batch sequences

  768

  Batch tokens

  201326592

  Steps

  1000

  Source: model-card · README.md; Training Details, Pre-training

  Its 201-billion-token budget corresponds to 1000 optimization steps, again without additional warmup.

  One third OCR'd PDFs plus length-reweighted mid-training data; no warmup, same constant LR; base weights are the average of its last 20 checkpoints. batch_tokens is 768 × 262,144 (card: 201.327 million).

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.

μP width transfer with spectral Muon (update scaled by √(d_out/d_in)); LM head and router LR ∝ 1/d; LM head zero-initialised

Width/depth parametrization used to transfer hyperparameters: SP (standard), μP, CompleteP, u-μP, depth-μP, with the multipliers that differ from SP. Only when stated.

Muon LR × batch grid (5e-4 to 8e-3; 4,608 or 6,912 sequences) on a half-width proxy (width 1280) trained 500B tokens, loss extrapolated to 20T with a power law in cumulative LR; picked LR 1e-3, batch 4,608; weight decay by the same extrapolation (2^-12 over 2^-13, 2^-14); fixed LR ratios across parameter groups

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.

MoE proxies M (7.8B total / 0.6B active, 97.6B tokens) and L (30.6B / 2.1B, 327.6B tokens) at 50 layers; 0.65B-active / 11.9B proxy pre-trained 2.79T tokens for mid-training mixes; thousands of 30M dense proxies (3B tokens) for the pre-training mix

Smaller models trained to choose the architecture or recipe (sizes, token budgets), as disclosed.

307.286

Calculated: train.tokens / params.total

train.tokens divided by params.total.

7,668

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.

20,801 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (1024 × 16/8) B/entry; SWA KV: 40 × (1024 × 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.

20,480 bytes

Calculated: 20480 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (1024 × 16/8) B/entry; SWA KV: 40 × (1024 × 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.

6,016 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (1024 × 16/8) B/entry; SWA KV: 40 × (1024 × 16/8) B/entry

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

788,524,039 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (1024 × 16/8) B/entry; SWA KV: 40 × (1024 × 16/8) B/entry

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

2,726,379,520 bytes

Calculated: 20480 B × 131072 (Global KV) + 81920 B × min(131072, 513) (SWA KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (1024 × 16/8) B/entry; SWA KV: 40 × (1024 × 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.