Ai2 Olmo 3
Ai2's fully open family at 7B and 32B with Base, Instruct, Think and RL Zero variants. Weights, training data, every intermediate checkpoint, logs and code under Apache 2.0.
Overview
Olmo 3 arrived 20 November 2025 from the Allen Institute for AI as dense models at 7B and 32B, each with four post-training variants. Base is the plain pretrained checkpoint. Instruct is the chat model. Think is the reasoning variant. RL Zero is the reinforcement learning starting point, released so other people can run their own RL experiments from a known state. Olmo 3.1 followed on 12 December 2025 with extended RL training, adding measurable gains on AIME, ZebraLogic, IFEval and IFBench. Olmo Hybrid, a 7B released 5 March 2026, replaces some attention layers with Gated DeltaNet and reaches Olmo 3's MMLU score on 49% fewer training tokens.
The reason to care is what ships alongside the weights. Ai2 calls it the model flow: the pretraining corpus, the mid-training and long-context data, the post-training mixtures, intermediate checkpoints across the whole run, training logs, the evaluation suite and the source for both the model and the surrounding infrastructure. All Apache 2.0. Most models described as open give you a weights file and a benchmark table. This gives you enough to re-run the work, which is a different category of artefact.
Be realistic about capability. Artificial Analysis puts Olmo 3.1 32B Think at 8 on its intelligence index, which is a long way below what a 31B Gemma or a 31.6B Nemotron does at the same memory footprint, and the 65,536 token context is tight for agent scaffolds. Where it earns its place is auditability, teaching and research: studying training dynamics, running controlled RL experiments, tracing outputs back to source documents through OLMoTrace, or answering a client who wants to know what the model was trained on before it goes near their data.
Key Features
- ✓ Dense 7B and 32B, each with four post-training variants: Base, Instruct, Think and RL Zero
- ✓ Apache 2.0 across the whole chain, not just weights: Dolma pretraining corpus, mid-training and long-context data, post-training mixtures, every intermediate checkpoint, training logs and the training code
- ✓ 65,536 token context on the 32B Think
- ✓ OLMoTrace maps a generated span back to matching documents in the training data, live in the Ai2 Playground
- ✓ Olmo Hybrid 7B, released 5 March 2026, swaps some attention layers for Gated DeltaNet and matches Olmo 3 on MMLU using 49% fewer training tokens
- ✓ olmo-eval, the evaluation harness that succeeded OLMES, released as a separate open component
Where it holds
- • Nothing else at this scale lets you inspect the corpus, so it is the only model here where a provenance question has a real answer instead of a model card
- • Olmo 3.1, released 12 December 2025, added 5-plus points on AIME, 4-plus on IFEval and 20-plus on IFBench over the launch checkpoint, showing the RL work is continuing rather than the family being parked
- • Apache 2.0 with no acceptable use policy, no user threshold and no geographic carve-outs, which matters for fine-tuning and redistribution
- • Intermediate checkpoints make it usable as a research substrate: you can study training dynamics or restart from a mid-training point, which frontier vendors never allow
Where it breaks
- • Capability is the trade-off. Artificial Analysis rates Olmo 3.1 32B Think at 8 on its intelligence index, well behind Gemma 4 31B and further behind the current open Chinese releases at similar size
- • 65K context is short by 2026 standards when Nemotron and Muse ship 1M and 131K respectively
- • Text only in the main line. Vision lives in Ai2's separate Molmo family, so this is not a drop-in multimodal option
- • Thin hosted-provider ecosystem, so if you want managed inference the choices are limited compared to more popular open families
My Take
Judge this one on what it provides that nothing else does, because on raw capability it loses. An intelligence index of 8 for the 32B Think means it is not going behind a customer-facing agent when Gemma 4 31B is available for the same memory budget. What you get instead is the entire chain: the Dolma corpus, mid-training and post-training mixtures, every intermediate checkpoint, the training logs and OLMoTrace to walk an output back to the documents that produced it. For a technical lead who has had to answer a regulator, a client or a procurement questionnaire about training data, that is the only credible answer on this list, and it is worth keeping one in the stack for that reason alone.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Free. Apache 2.0 downloads on Hugging Face covering weights, the Dolma pretraining corpus, post-training mixtures and intermediate checkpoints. The Ai2 Playground is free to use in the browser. If you want an endpoint, Olmo 3.1 32B Think is $0.15 per million input and $0.50 per million output on OpenRouter, checked 28 August 2026. Running the 32B yourself needs roughly a 48GB card at BF16 or a 24GB card quantised.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Ai2 Olmo 3 holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Ai2 Olmo 3
Google Gemma 4
open-sourceGoogle's open-weight family in five sizes, 2B effective up to 31B dense, text and image in with audio on the smaller ones, 256K context and Apache 2.0 weights since April 2026.
NVIDIA Nemotron 3
open-sourceNVIDIA's open-weight family, 31.6B Nano up to the 550B Ultra MoE. Hybrid Mamba-Transformer, 1M context, weights plus roughly 3T tokens of training data published.
Mistral AI
open-sourceEuropean lab publishing open weight models you can run on your own hardware. Large 3 is a 675B MoE with 41B active, 256K context and Apache 2.0 weights.
InternLM
open-sourceShanghai AI Lab's open-weight family (书生·浦语), now the Intern-S scientific multimodal line: 35B up to 1T parameters, Apache 2.0, trained on domestic computing infrastructure.