Foundation, Frontier and Open-Weight Models
Three overlapping labels for large models:
- Foundation model: a model trained on broad data at scale that can be adapted to many downstream tasks (term introduced by Stanford’s Center for Research on Foundation Models in August 2021). LLMs, image and multimodal models all qualify.
- Frontier model: one at or near the leading edge of capability; the term is used in safety frameworks and regulation, where extra obligations attach to the most capable or compute-intensive models (ai-safety-and-governance, ai-regulation-and-policy).
- Open-weight model: weights are downloadable and runnable locally, under a licence that may still restrict use. “Open weights” is not the same as open source, which would also need training code, data and an open licence (open-source-ai).
Where to look in the vault
Model families: anthropic-claude, gpt-5-series, gemini-3-series, deepseek-family, kimi-k2-series, glm-5-series, mistral-family, gemma-4. Labs and companies: ai-research-labs, openai, anthropic, meta, mistral-ai, deepseek, nvidia. Comparison tools: ai-benchmarks-and-evals.
Related: large-language-model, scaling-laws, sovereign-ai.
Sources
- https://en.wikipedia.org/wiki/Foundation_model (secondary; term origin and definition; re-fetched 2026-10-07: introduced by Stanford HAI CRFM, August 2021)
- https://en.wikipedia.org/wiki/AI_alignment (secondary)
Open items
- ‘Frontier model’ and ‘open-weight vs open source’ definitions are stable usage, not checked against a primary regulatory text here.
- The AI_alignment Wikipedia source is unrelated to the claims above and was not re-checked.
- Regulatory thresholds for ‘frontier’ (compute or capability) change; check ai-regulation-and-policy for current status.