Falcon family (7B/40B, Falcon 3, Falcon-H1 / H1R)

Falcon 7B/40B, the Falcon 3 line, and the current anchor Falcon-H1R-7B are all downloadable from the tiiuae Hugging Face organization.

Terms vary materially by version: early releases were plain Apache 2.0, while Falcon 3 and Falcon-H1/H1R use the TII Falcon LLM License 2.0, an Apache-derived license with an Acceptable Use Policy, mandatory attribution notice, no-litigation clause, and commercial obligations, so openness must be checked at the exact release.

Independently deployable from a single 7B checkpoint up to the 180B model with standard open serving stacks; no vendor runtime is required.

Model cards publish dataset composition for several releases, including RefinedWeb proportions, and the H1R technical report documents a two-stage pipeline, the SFT data mix, and a difficulty-aware weighting scheme, but the training corpora themselves are not published.

Better documented than most weight-only releases, with published methodology and hardware detail, but not uniformly reproducible across the family.

TII ships Falcon weights and serving examples under a central repository that accepts a trickle of outside pull requests, but the training stack, dataset weighting, and release cadence are set by the institute alone.