Olmo-3-1125-32B

2025-11-20

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").

32B parameters

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

32B 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.

32.23B parameters

Calculated: embedding + output head + all experts + other weights, allenai/Olmo-3-1125-32B@c2b61da 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.

1.027B parameters

Calculated: input embedding + output head, allenai/Olmo-3-1125-32B@c2b61da safetensors headers

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

31.21B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, allenai/Olmo-3-1125-32B@c2b61da 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.

487.6M parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, allenai/Olmo-3-1125-32B@c2b61da 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.

31.21B parameters

Calculated: all weights except embedding and output head, allenai/Olmo-3-1125-32B@c2b61da 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.

64

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

5,120

Residual stream width.

100,278

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

No

Input embedding and output head share weights.

OLMo 2 tokenizer, derived from OpenAI cl100k

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

80

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).

3 SWA : 1 global (last layer 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.

16

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

3

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').

48

Layers with a fixed attention window.

40

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.

5,120

Calculated: heads × head_dim

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

5

Calculated: heads / kv_heads

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

128

Inferred from the source; see the explanation below.

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

4.096K tokens

Sliding-window size of the local layers.

8

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

40

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

128

Inferred from the source; see the explanation below.

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

RoPE

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

500,000

RoPE base frequency of the global layers at release.

500,000

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, ...).

27,648

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

No

FFN projections have bias terms.

5.4

Calculated: ffn.dim / d_model

ffn.dim divided by shape.d_model.

0

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

RMSNorm

Normalization layer type in the blocks.

post (on sublayer outputs, inside the residual branch)

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.

65.54K tokens

Longest sequence length trained on after context extension.

65.54K tokens

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

YaRN on full-attention layers only (SWA RoPE unchanged); 8K → 65K over 100B tokens after midtraining, 34% long / 66% midtraining data, best-fit packing, intra-document masking

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

5.7T tokens

Inferred from the source; see the explanation below.

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

AdamW (OLMo-core SkipStepAdamW)

Optimizer used for the bulk of parameters.

0.0006

Peak learning rate of the main pretraining stage.

6.21e-05

Learning rate at the end of the decay.

2000-step warmup; cosine over one epoch (5.93T) targeting 10% of peak, truncated at 5.5T; midtraining and long-context: linear decay 2.071e-4 → 0

Learning-rate schedule shape with phase boundaries.

2000 steps

Warmup length in steps or tokens.

1,024 × 8,192 (8.39M tokens) pretraining; 512 × 8,192 (4.19M tokens) midtraining; 128 × 65,536 (8.39M tokens) long-context

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

8.389M tokens

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

0.1

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.

z-loss 1e-5; QK-norm; no weight decay on embeddings

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

intra-document masking during long-context extension

Whether attention crosses document boundaries within a packed sequence.

Dolma 3 Mix, 5.93T: Common Crawl 76.1%, olmOCR science PDFs 13.6%, Stack-Edu code 6.89%, FineMath 3+ 2.56%, arXiv 0.86%, Wikipedia/Wikibooks 0.04%; upsampled from a 9.31T pool

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

December 2024

Pretraining data cutoff date.

H100 80GB; pretraining 512 GPUs for ~9.5 days then 1024 for ~35 days; midtraining 2 × 512; long-context 1024

Accelerators used for pretraining, count and duration where stated.

1,024 H100

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").

HSDP (DP-shard 64 · DP-rep 16)

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").

midtraining: weights of two 100B runs with different data-order seeds averaged; long-context: checkpoints at steps 10,000, 11,000, 11,921 averaged

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

• pretraining

  Tokens

  5500000000000

  Seq len

  8192

  Lr

  0.0006

  Batch tokens

  8388608

  Source: paper · p.3 §2; p.30–35 §3.5–3.6; p.85 Table 35

  The initial pretraining phase on 5.5T tokens took about 9.5 days on 512 GPUs, followed by an additional 35 days on 1024 GPUs.

  Dolma 3 Mix; cosine to 5.93T truncated at 5.5T.

• midtraining

  Tokens

  100000000000

  Seq len

  8192

  Lr

  0.0002071

  Batch tokens

  4194304

  Source: paper · p.3 §2; p.30–35 §3.5–3.6; p.85 Table 35

  Midtraining consisted of two parallel runs on 512 GPUs each, covering 100B tokens per run

  Dolma 3 Dolmino Mix; no warmup, linear decay to 0; two runs with different data-order seeds, weights averaged.

• long-context

  Tokens

  100000000000

  Seq len

  65536

  Lr

  0.0002071

  Batch tokens

  8388608

  Steps

  11921

  Source: paper · p.3 §2; p.30–35 §3.5–3.6; p.85 Table 35

  Long-context extension was executed as a single run on 1024 GPUs

  Dolma 3 Longmino Mix; 200-step warmup, linear decay to 0; final model averages checkpoints at steps 10,000, 11,000, 11,921.

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.

Data ablations up to 1B params at 100B tokens; mix swarms of 30M models at 3B tokens (5× Chinchilla)

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

178.125

Calculated: train.tokens / params.total

train.tokens divided by params.total.

182.653

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.

0.0245281

Calculated: train.batch_tokens / (train.lr_peak × train.weight_decay × train.tokens)

B / (η λ D) with B = train.batch_tokens, η = train.lr_peak, λ = train.weight_decay and D = train.tokens: the AdamW EMA timescale as a fraction of training (Bergsma et al. 2025, Power Lines, arXiv 2505.13738). Computed only when the optimizer is AdamW; uses the peak LR.

71,680 bytes

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

65,536 bytes

Calculated: 65536 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 16 × (2048 × 16/8) B/entry; SWA 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.

2,240 bytes

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

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

293,601,280 bytes

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

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

9,395,240,960 bytes

Calculated: 65536 B × 131072 (Global KV) + 196608 B × min(131072, 4096) (SWA KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 16 × (2048 × 16/8) B/entry; SWA 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.

6

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).