Olmo 3 32B
Ai2 · OLMo · open weights
- GQA
- Sliding window
- RoPE
- Post-norm
- QK-norm
Facts and where they come from
| Released | 2025-11 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | apache-2.0 | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 32B | labmodel cardmodel name Olmo-3-32B |
| Active parameters | not disclosed | not disclosed |
| Context length | 64K tokens | config.jsonconfig.jsonmax_position_embeddings |
| Norm placement | post | codemodelling codetransformers 5.18.0 olmo3: post_attention_layernorm and post_feedforward_layernorm only |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 olmo3: post_attention_layernorm and post_feedforward_layernorm only |
| QK-norm | yes | codemodelling codetransformers 5.18.0 olmo3: q_norm and k_norm |
| Positional encoding | RoPE | codemodelling coderotary on the full head (default) |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 olmo3: post_attention_layernorm and post_feedforward_layernorm only |
Architecture, drawn from the data
48× GQA 40q/8kv, window 4096 + 16× GQA 40q/8kv. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.
Modelled costs
From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).
| Parameters (modelled) | 32.2B |
|---|---|
| Active per token (modelled) | 32.2B |
| Without embeddings and output head | 31.2B total, 31.2B active |
| Published weights (Hugging Face count) | 32.2B |
| KV cache per token, BF16 (layers that grow with context) | 64 KiB |
| KV cache + state at 64K tokens, BF16 | 4.75 GiB |
| Decode FLOPs per token at 4K context | 68.8 GFLOP |
| Prefill FLOPs for a 4K prompt | 267 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 5,120 | config.jsonconfig.jsonhidden_size |
| vocab | 100,278 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 40 | config.jsonconfig.jsonnum_attention_heads |
| mixers.full.kv_heads | 8 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 128 | codemodelling codetransformers 5.18.0: head_dim = hidden_size / num_attention_heads |
| mixers.full.qk_norm | true | codemodelling codetransformers 5.18.0 olmo3: q_norm and k_norm |
| mixers.sliding.type | attn | codemodelling codeattention block |
| mixers.sliding.heads | 40 | config.jsonconfig.jsonnum_attention_heads |
| mixers.sliding.kv_heads | 8 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.sliding.head_dim | 128 | codemodelling codetransformers 5.18.0: head_dim = hidden_size / num_attention_heads |
| mixers.sliding.window | 4,096 | config.jsonconfig.jsonsliding_window |
| mixers.sliding.qk_norm | true | codemodelling codetransformers 5.18.0 olmo3: q_norm and k_norm |
| ffns.dense.type | dense | codemodelling codeMLP block |
| ffns.dense.d_ff | 27,648 | config.jsonconfig.jsonintermediate_size |
| ffns.dense.gated | true | codemodelling codeMLP: gated (SwiGLU/GeGLU) |
| layout | 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense · 3× sliding/dense · 1× full/dense | config.jsonconfig.jsonlayer_types |
Sources
- config.json @ f2edda1
- model card
- arXiv 2512.13961
- modelling code · transformers 5.18.0 modelling code, or the model repository's own modelling file at the pinned revision
Listed in the LLM Architecture Gallery checklist as “OLMo 3 (32B)” (name only; see about).