Open Weights
AI models whose trained parameters are publicly released for download and local deployment — enabling privacy, customization, and offline use — but whose training data, code, and full methodology may not be disclosed.
What it means
An open-weights model is one whose trained parameters (weights) are publicly released and can be downloaded, run, and fine-tuned without going through a company’s API. This is distinct from fully open-source models, where training code, data, and infrastructure are also disclosed.
Key open-weights models relevant to research:
- Meta Llama 3 — general-purpose language model, commercially usable with Meta’s license
- Mistral — series of capable smaller models from Mistral AI (Paris)
- Falcon — from TII Abu Dhabi, commercially permissive license
- ESM3 — protein sequence and structure model from EvolutionaryScale, open weights with academic/commercial tiers
Open weights ≠ open source: A model can release its weights while keeping the training data, training code, and full methodology proprietary. The phrase “open source AI” is used inconsistently in the industry — check specifically whether weights, training data, and training code are released when evaluating openness claims.
Why it matters for researchers
Privacy and data sensitivity: Running an open-weights model locally means your prompts and data never leave your institution. For research involving unpublished data, patient records, or proprietary chemical structures, this can be critical.
Reproducibility: Open weights allow exact replication — if you publish research using a model and share the weights, others can reproduce your results precisely, unlike API-based models that change versions over time.
Fine-tuning and specialization: Open-weights models can be fine-tuned on domain-specific data to improve performance on specialized tasks (clinical note extraction, chemical property prediction, etc.).
The practical tradeoffs:
- Open-weights models generally require significant compute to run (GPUs, server infrastructure)
- Smaller open-weights models often underperform larger proprietary models on general tasks
- Specialized domain models (like ESM3 for proteins) may outperform general models for specific tasks even at smaller scale