OLMo model series: every generation compared

OLMo is Ai2's fully open language model line. Compare Olmo 1, Olmo 2 and Olmo 3 by size, context length and license, plus OLMoE, olmOCR and Molmo.

What OLMo is and who makes it

OLMo is a family of language models from Ai2 (the Allen Institute for AI). Its defining trait is openness beyond the weights. The original OLMo paper describes a release of model weights, open training data and training and evaluation code, and the later generations keep that pattern: the Olmo 2 paper lists weights, full training data, training code and recipes, training logs and intermediate checkpoints. For a developer that means you can inspect what a model was trained on, reproduce a fine-tune from a known checkpoint, and pick a license without surprises: the cards we checked for every generation below say Apache 2.0.

The generations in order

Olmo 1 (February 2024)

Olmo 1 is the first release, described in a paper submitted in February 2024. The 7B model card lists 2.5 trillion training tokens, a context length of 2,048 tokens, full attention (not grouped query attention), SwiGLU activations and rotary position embeddings, under Apache 2.0. Two thousand tokens of context is short by current standards, so this generation is mostly of interest for research and for studying training dynamics from its published checkpoints.

Olmo 2 (November to December 2024)

Olmo 2 is a set of dense autoregressive models. The paper (submitted December 31, 2024) presents 7B, 13B and 32B sizes, with a modified architecture, training stability changes and a new data mixture for the final training stage. The 7B card lists 4 trillion training tokens and the 13B card 5 trillion, with a context length of 4,096 tokens, under Apache 2.0. Post-training uses supervised fine-tuning, DPO and PPO. Context is still short, but the 7B and 13B sizes are easy to host and fine-tune.

Olmo 3 (November 2025)

Olmo 3 was announced on November 20, 2025, with an Olmo 3.1 update on December 12, 2025. Ai2's announcement lists 7B and 32B sizes, and four tracks per size: Base, Think (reasoning with intermediate thinking traces), Instruct (chat and tool use) and RL Zero (an open reinforcement learning path for bootstrapping reasoning). The announcement states support for up to roughly 65K tokens of context, a jump from the 4,096 of Olmo 2, and describes a three-stage pretraining recipe over a new corpus called Dolma 3 of about 9.3 trillion tokens. The Olmo 3.1 32B Think card says the update is mainly extra reinforcement learning on the 32B Think model. Cards for the Olmo 3 models are Apache 2.0.

Which generation to use today

All three generations are Apache 2.0 per their cards, so the license does not decide the choice. Size, context length and task do:

  • Long documents or agent loops: Olmo 3, because it is the only generation with context in the tens of thousands of tokens.
  • Reasoning-style output: the Olmo 3 Think variants, which emit intermediate thinking. Use Instruct when you want direct chat answers and tool use.
  • Single mid-size GPU: the 7B Olmo 3 variants, or Olmo 2 at 7B or 13B if you already have a recipe built around them.
  • Largest open option from Ai2: the 32B sizes in Olmo 3 or Olmo 2.
  • Research on training itself: Olmo 1 and Olmo 2 expose checkpoints and data that make them useful for studying how a model changes during training.

Treat any benchmark claim as vendor-reported: Ai2's own pages are the source for the Olmo 3 and Olmo 3.1 results, so run your own evaluation on your task before committing.

Side lines

OLMo is also the name behind several related models, each with its own hub:

  • OLMoE is a mixture-of-experts model. OLMoE-1B-7B-0924 (September 2024) has 7B total parameters with about 1.3B active per token, Apache 2.0. Total parameters set the weight memory you must hold, while active parameters set the compute per token.
  • olmOCR is a document OCR model. olmOCR-2-7B-1025 is built on Qwen2.5-VL-7B-Instruct and tuned with reinforcement learning for PDFs with equations and tables, Apache 2.0. It is a vision-language model, not a general chat model.
  • Molmo is a vision-language line. Molmo-7B-D-0924 pairs a Qwen2-7B language model with a CLIP ViT-Large vision encoder and is Apache 2.0. Note that it does not use an OLMo backbone.

Running it

Every model above can run on rented GPUs through Aquanode GPUs; pick the size from the generation list above and check the model's hub for memory estimates.

OLMo generations

Every OLMo generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.

HubModelsSizesSmallest native VRAMCheapest live fit for itEst. $/hr
OLMo 121.2B to 6.9B5.3 GBRTX 3070$0.088/hr
OLMo 251.4B to 32.2B3.3 GBRTX 3070$0.088/hr
Olmo 387.3B to 32.2B16.3 GBRTX A5000$0.176/hr

Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.

Other OLMo lines

Side lines from Ai2: specialised or companion models outside the main OLMo generations.

HubModelsSizesSmallest native VRAMCheapest live fit for itEst. $/hr
OLMoE46.9B15.5 GBRTX A4000$0.167/hr
olmOCR28.3B to 8.3B18.5 GBRTX A5000$0.176/hr
Molmo34.9B to 8.7B21.7 GBRTX A5000$0.176/hr

Run and fine-tune OLMo

Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.

Best models by task

OLMo models appear on these ranked task pages.

Sources

Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.

Start ComfyUI on a rented GPU.

© 2026 Aquanode. All rights reserved.

All trademarks, logos and brand names are the property of their respective owners.