Ai2

A hybrid attention + GDN (gated DeltaNet) model. See our blog post for more insights about the architecture and its challenges.

Qwen3 fine-tuned coding models, showing that you can train highly performant models for a codebase even without unit tests or verified data.

A VLM which combines Olmo 3 7B and SigLIP. As usual for Ai2 releases, the full data for the training is released.

As it turns out, letting a RL training run continue will make models better (an extra 3.5 weeks for this one)! 3.1 is an improvement on all metrics over its predecessors. The paper is also updated and dives even deeper into all the details.

A series of truly open models, covering everything from data to all models from the model flow. In case you somehow missed it, check out the coverage of the model: Olmo 3: America's truly open reasoning models

An earth observation model.

A version of Molmo for robotic tasks.

An open reproduction of the Whisper models, coming in multiple sizes.

A new model by Ai2, where different organizations can train experts on their data to improve a shared model. The blog provides more details about the model and its training process.

A robotic model based on Molmo.

An update to the OLMo 2 models.

A truly open model release, rivaling the quality of GPT-3.5 and the OG GPT-4. For more information, see our post. The base models are released alongside the instruction models.

Ai2 released an updated version of Tülu. Trained on the same data as the previous version, but with GRPO (instead of PPO), the same algorithm used by R1. This results in better performance across the board, most notably in math benchmarks. It also is evidence against the argument that GRPO is "poor man's PPO". Full reasoning models from Ai2 are still "coming soon."

Ai2 has released an OCR model based on Qwen2-VL 7B. Alongside the model, which is released under Apache 2.0, they released a library supporting the entire pipeline - from training to evaluation to inference.

We updated our OLMo 7B model with some small improvements to training data and annealing techniques. A pretty minor bump, but we have some exciting new models coming soon. An interesting thing I've learned about pretraining is how loss spikes often relate to "skipped tokens," making the models worse at a fixed compute budget. This is because the gradients is loss spikes get clipped and the data effectively does nothing.

A Llama 2 fine-tune designed to specialized on scientific information extraction and processing tasks. Built on top of our Tulu 2 work!