Ai2 cracked open every drawer in the AI cabinet, here's what's inside

The Allen Institute for AI (Ai2) has unveiled Olmo 3, a family of fully open language models. The pitch: complete transparency from data to deployment. Plenty of companies have made that claim before, but few have backed it up. Olmo 3 opens every drawer in the cabinet. The raw pretraining mixtures, the mid-training refinements, the post-training instruction sets, the evaluation harnesses, and even the tooling used to clean and deduplicate data before training began. All of it is public. Microsoft just opened a codebase that kills the worst…

Olmo 3 comes in six variants at two parameter scales, 7B and 32B, each in three flavors: base, instruct, and think. The 32B-Think variant does chain-of-thought reasoning, surfacing intermediate steps for complex prompts. The 32B-Instruct is Ai2's most capable fully open chat model to date, with tool use and multi-turn dialogue support. The 7B variants aim for efficiency: they run on a broader range of hardware while keeping performance competitive. There are benchmark scores on the model cards, and they look solid for programming, reading comprehension, and math. But the real story is what Ai2 calls the 'model flow.' This is the full lifecycle: data curation through pretraining, mid-training, long context adaptation, instruction tuning, preference optimization, reinforcement learning, and finally the reasoning-specialized 'think' branch. Sipp.sh Launches Open-Source Library for Local AI…

A transparent data pipeline

Ai2 provides download links for the pretraining data mixture, a fully open blend of curated web text, code, books, and scientific articles, deduplicated and quality-filtered. Mid-training data, used to refine the base model with domain-focused mixtures, is also available. Post-training data covers supervised instruction responses and comparison data used in direct preference optimization (DPO) and reinforcement learning stages.

The tooling is equally open. The training framework, OlmoCore, is available for fast configuration. Data preprocessing tools include Duplodocus for ultra-efficient fuzzy deduplication and Datamap-rs for large-scale cleaning. The post-training pipeline, Open Instruct, lets researchers replicate or modify the instruction tuning process. For evaluation, Ai2 offers OLMES for reproducible evals and Decon to help remove test set contamination from training data. Hugging Face's Moon Bot turns Slack into a coding…

OlmoTrace, possibly the most interesting piece, lets users trace model outputs back to specific training data points. That could prove critical for diagnosing hallucinations or bias. It also matters for AI governance research that requires provenance at the example level.

Why fully open matters now

The release lands at a moment when the definition of 'open' in AI has become deeply contested. Meta's Llama series, widely described as open, ships under a custom license that restricts commercial use for large-scale applications. Mistral's models carry similar restrictions. Even models labeled 'open source' by their creators often withhold training data, training logs, and intermediate checkpoints.

Olmo 3 sidesteps those compromises. The entire flow, every checkpoint, every decision about data filtering, every hyperparameter, is public. This is not a concession. It is the point. Ai2 treats transparency as a prerequisite for scientific rigor: if a result cannot be reproduced and the data behind it cannot be examined, the field is operating on trust rather than evidence. DeepSeek's DSpark just fixed the two things that held…

Forbes journalist Janakiram MSV described Olmo as 'stand(ing) out by providing full access' compared to current open LLMs. Hugging Face CEO Clem Delangue said Olmo 'ensures complete transparency and sets a strong foundation for transformative work.'

Already generating research

The transparency is already paying dividends. Ai2 highlights three research projects building on earlier Olmo checkpoints: machine unlearning research using Olmo-7B as a testbed for removing specific data influence without retraining; clinical NLP applications for analyzing medical text while preserving data transparency; and fundamental studies into learning dynamics and scaling behaviors enabled by access to training logs and intermediate checkpoints. Anthropic Launches Claude Science, an AI Workbench…

Simon Mo, project co-lead at vLLM, noted that Olmo's architecture pushes 'the frontier of open-source model design.' Anastasios Angelopoulos, CEO of LMArena, described Olmo as 'becoming the instrument through which the community builds the next layer of open, foundational intelligence.'

What is not yet known

Ai2 has not released detailed benchmark scores against comparable models from Meta, Mistral, or Qwen. No latency figures or inference cost comparisons are published either. The model flow page does not specify the hardware used for training, the total compute budget, or the power consumption. The Think variant's reasoning steps are surfaced, but the company has not described the chain-of-thought methodology in detail. These gaps are notable for a project built on transparency. But the data and code are public, meaning independent audits can fill them. Ifbench: the new benchmark testing AI instruction following

The company is accepting subscriptions for monthly updates about Ai2's work, suggesting Olmo 3 is not a one-off release but part of a sustained effort to redefine openness in AI.