Allen Institute for AI, or Ai2 for short, has just introduced Olmo 3, a family of artificial intelligence models that differs from most others. While many companies guard their training processes like a chef guards the recipe for their best dish, Ai2 shares everything. Olmo 3 is not just another "open" model that offers only weights—the numbers that determine its responses. Here, you get everything: the weights, every saved training snapshot, data-related decisions, and every line of code that transformed raw data into a thinking machine. It is as if Tesla shared not only its car designs but the entire factory.
This model is the first fully open reasoning model with 32 billion parameters. Parameters are adjustable values; the more there are, the smarter the responses. Reasoning means that the model shows its thought process step by step. You can trace every response back to the exact training data, modify it, rebuild it, or improve it. The model family ranges from 7 billion to 32 billion parameters and runs on everything from laptops to data centers.
Olmo 3 models and their strengths
Olmo 3-Think, with 32 billion parameters, is a frontier reasoning model that shows its thinking step by step and is fully open for research and reinforcement learning. It is ideal for complex tasks where you need to see how the model reached its conclusion. Olmo 3-Base, available with 7 billion and 32 billion parameters, excels at programming, reading comprehension, and mathematics. It has a context window of 65,000 tokens, which corresponds to approximately 50,000 words. This means it can analyze long texts, such as entire books or complex documents.
Olmo 3-Instruct, with 7 billion parameters, is optimized for conversations and tool use. It responds quickly, handles multi-turn questions, and works well in chat applications. Its efficiency is remarkable: Olmo 3-Think competes with similarly sized models from Qwen but was trained on 6 times fewer tokens. This means lower costs, less energy, and faster iteration without affecting performance.
Training costs were about $500,000 for the 7-billion-parameter model, which is approximately CZK 10,485,000, and around $2.2 million for the 32-billion-parameter model, or about CZK 46,134,000. That is a fraction of what is spent on closed models, and everything is available for free.
The significance of Olmo 3
For a long time, "open source" in AI meant only releasing the weights, while training secrets remained locked away. With Olmo 3, researchers can reproduce results, companies can customize models without vendor dependency, and developers can understand model behavior. Large technology companies want others to depend on their opaque models, but Ai2 provides access to the entire model.
The models are available on Hugging Face and in the Ai2 Playground under the Apache 2.0 license, which allows free use, modification, and commercial use. The code is on GitHub, and full technical details are available. It fully understands Czech.
Olmo 3 is built on Dolma 3, a new corpus containing approximately 9.3 trillion tokens from websites, scientific PDFs, code, mathematical problems, and encyclopedic text. For the mid-training phase, they used Dolma 3 Dolmino, a mixture of 100 billion tokens focused on mathematics, science, code, and reading comprehension. For long context, they added Dolma 3 Longmino with 50 billion tokens from long documents.
Performance and comparison with competitors
Olmo 3-Base 32B outperforms other fully open base models such as Marin 32B and Apertus 70B in mathematics, programming, and reading. For example, it achieved 80.5% on GSM8k, 43.4% on MATH, and 43.9% on BigCodeBench. It also competes with models such as Qwen 2.5 32B and Gemma 3 27B.
Olmo 3-Think 32B is the strongest fully open reasoning model, approaching Qwen 3 32B with 96.1% on MATH and 89.8% on BigBenchHard. It achieved 91.4% on HumanEvalPlus. Olmo 3-Instruct 7B outperforms Qwen 3 8B in safety with 87.3% and is strong in tool use, achieving 79.3% on SimpleQA.
Training was performed on a cluster of up to 1,024 H100 GPUs, with a throughput of 7,700 tokens per second per device for the 7B model. They accelerated post-training by 8 times for SFT and 4 times for RL through improvements such as continuous batching and on-the-fly weight updates.
Olmo 3 integrates OlmoTrace, a tool for tracing outputs back to training data in real time. This makes it possible to see why the model responds in certain ways and adjust the data as needed.
For post-training, they introduced Dolci, a dataset suite for reasoning, tool use, and instruction following. It includes mixtures for SFT, DPO, and RLVR. Olmo 3-RL Zero 7B provides a path for reinforcement learning directly on the base model, with checkpoints for mathematics, code, instructions, and chat.
The models run on a single GPU with 80 GB of memory for the 32B version, making it ideal for research. Ai2 plans further improvements, such as mixtures of experts and better character-level training, but Olmo 3 is already setting the standard for openness.
You can chat with the model here: playground.allenai.org
Additional source: interconnects.ai
