tools
OLMo
OLMo is Ai2’s fully open language model family with public weights, data, code, checkpoints, and evaluations.
OLMo provides language models for research, development, and local deployment. Its model flow includes pretraining, mid-training, post-training, datasets, checkpoints, training code, and evaluation tools.
The current product page presents the Olmo 3 family in 7B and 32B Base, Think, and Instruct variants. Ai2 provides artifacts for download and documentation for running models locally; hosted usage pricing was not found.
Features
- Olmo 3 models in 7B and 32B Base, Think, and Instruct variants
- Public model weights, training data, checkpoints, code, and evaluations
- Supports programming, reading comprehension, mathematics, and extended-context tasks
- Think variants provide reasoning-focused checkpoints
- Instruct variants support chat, tool use, and multi-turn dialogue
- OlmoCore provides a framework for language-model training
- OLMES provides tools for reproducible model evaluation
- Open Instruct provides a post-training pipeline
Use cases
- Run language models locally on available hardware
- Fine-tune base checkpoints for specialized applications
- Study how training data and methods affect model behavior
- Build chat and tool-use systems with instruction-tuned checkpoints
- Reproduce or extend language-model training experiments
Pros
Cons
Latest updates
- archive/speed-process-workers
Thread prefetch remains the default for mixed mid-training; process mode is available.
- archive/speed-norms-interval
Gradient-norm bookkeeping now runs once per step.
- archive/speed-no-locks
OLMoOCR pages are rendered in helper processes using thread-local shard handles.
- archive/speed-full-window (OLMo 3.5)
The OLMo 3.5 recipes leave prefetch_keep_full off.
- archive/speed-device-once: Stage the pixels to the device through a pinned buffer on a side stream
Image batches are staged through a reusable page-locked buffer and copied on a side stream.
Capabilities
- Self-hosted — “Olmo 3 is a family of compact, dense models at 7 billion and 32 billion parameters that can run on everything from laptops to research clusters.” source
- API — “or use them via API courtesy of our inference partners.” source
- Runs models for you — “or use them via API courtesy of our inference partners.” source
Get it
Pricing
- Prices checked
- 2026-10-25