Falcon 40B Instruct
Falcon 40B Instruct is an early 40B chat model adapted from Falcon-40B. The downloadable release uses Apache 2.0 and remains available as a historical Falcon checkpoint.
Training and release background
Its instruction training used about 150 million Baize tokens mixed with 5% RefinedWeb data. The model uses rotary position embeddings and multiquery attention. The card recommends the separate base release when building a new instruction model.Falcon 40B Instruct model card
Separate historical reproduction from model choice
Practical guidance. Use this checkpoint when its exact training lineage matters to the work. If the task is general writing or extraction, compare it with newer models at the required latency. A larger parameter count by itself is not a reason to prefer an older release.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/falcon-40b-instruct |
| License | Apache 2.0Falcon 40B Instruct model card |
| Release status | Earlier release; published weight access remains availableFalcon 40B Instruct model card |
Related models
Falcon 180B Chat is a 180B text model adapted from Falcon-180B.
Launch date not confirmedRestricted accessDownloadableFalcon 7B Instruct is an early Falcon chat and instruction model based on Falcon-7B.
Launch date not confirmedDownloadable