Olmo 3 7B Instruct
Fully open 7B instruction model with published checkpoints, training code, data, and evaluation artifacts. Its open release supports independent inspection and self-hosted privacy controls.
Score calculation
(7 + 7 + 9) ÷ 3 = 7.7
3 of six dimensions rated. Unrated dimensions and BDS status are excluded.
Full scoring method ↗Setup details
Olmo 3 source/training statements refer to the family and model flow. Energy, labour and use relationships are not implied by openness; remaining claims are limited to published project materials.
AI2's release documentation identifies Olmo 3 7B Instruct as part of its November 2025 release and publishes model-flow artifacts. Checked 2026-10-23; no newer Olmo language-model release was verified in the official docs used.
Energy
↑AI2's official OLMo 3 training README reports 5.93 trillion pretraining tokens for the 7B base model and identifies the 7B Instruct post-training stages/checkpoints. Training scripts disclose accelerator counts for portions of the 7B pipeline, but the checked README/card do not give end-to-end GPU-hours, device energy, facility electricity, emissions, or water. These base-family details are not a complete measured footprint for the final Instruct checkpoint; no standardized inference-energy figure is reported.
Reasoning & sources
AI2 discloses token volume and parts of the training pipeline, but not end-to-end energy, emissions, or inference use for the final Instruct checkpoint. Those details make the run more reproducible without quantifying its footprint; compare only after a bounded energy report exists.
No complete 7B Instruct pipeline runtime/energy/emissions or standardized inference measurement; accelerator counts alone do not establish electricity use.
Training rights
↑AI2 publishes the Dolma 3/Dolci data curriculum and says it follows web standards, excluding sites that explicitly disallow crawling and paywalled content. This is meaningful provenance disclosure but not a complete legal clearance of every source.
Reasoning & sources
AI2 publishes Dolma 3/Dolci sources and describes excluding sites that disallow crawling or paywalled content. That is a meaningful rights-respecting process, though web crawling rules do not independently settle every copyright claim or source permission.
Published corpus and rights-handling disclosures are unusually detailed, but remain developer-reported and lack independent rights verification; score reflects transparency with that material limitation.
No independent rights audit; dataset curation and rights status remain developer-reported.
Privacy
↑Weights can be run locally, letting an operator keep inference inputs on their own infrastructure; hosted deployment privacy depends on the operator. This does not establish privacy properties of AI2-hosted demos.
Reasoning & sources
AI2 offers local weights and deployment guidance, allowing an operator to keep prompts on its own infrastructure. The materials do not address local telemetry or hosted vendors, so privacy depends on the chosen deployment.
Public weights enable local deployment and operator control over prompts, but that is a deployment option rather than a guaranteed privacy safeguard; hosted services and downstream configurations require separate review.
No complete hosted-demo retention policy assessed.
Openness
↑AI2 publishes weights, code, pretraining and post-training datasets, and evaluation methods, and describes the family as fully open. This supports strong reproducibility compared with model-only weight releases.
Reasoning & sources
AI2 releases weights, code, training datasets, and evaluation methods, enabling substantial independent inspection and reproduction. Reproducing a frontier-scale run still requires substantial resources, and some operational dependencies may not be captured in the released materials.
Weights, training data, code, evaluation methods, and training stages are published.
Check exact checkpoint license and dependencies for a specific deployment.
Worker treatment
↑The release describes human preference data in post-training but does not disclose worker count, pay, working conditions, or contractor/vendor chain for Olmo 3.
Reasoning & sources
The Olmo 3 release identifies human preference data but does not disclose annotator pay, working conditions, or vendor arrangements. That confirms human input while leaving the fairness of its production unknown; request worker sourcing details.
Human data contribution is described; labour conditions remain undisclosed in the source checked.
Government & military
↑Ai2 publishes Olmo-3-7B-Instruct under Apache 2.0 and provides the checkpoint for independent use. That enables downstream deployment without an Ai2 service contract; no cited source establishes military use.
Reasoning & sources
The exact model card gives Apache 2.0 terms, so a third party can run the checkpoint independently of Ai2. This limits what can be inferred from company contracts or hosted-customer disclosures; it does not show that any military organization has adopted Olmo.
No complete downstream-use register for open weights.
BDS status
↑We did not find AI2 (Allen Institute for AI) among the named targets in the two official BNC campaign lists. This is a limited list check, not BDS approval.
Reasoning & sources
AI2 (Allen Institute for AI) is not a named target in the checked lists, but BDS uses selective campaigns, so absence cannot establish a clean record. Running downloaded weights locally is a different decision from paying the developer or a cloud host; check the host as well.
This review covers the named targets in the BNC consumer/pressure guide and No Tech campaign, not every BDS campaign, local group, investment or supplier. Lists and ownership can change. A company’s absence from these lists is not an endorsement or proof that it has no relevant relationships.