Falcon 40B

Falcon 40B is a large autoregressive text language model from the Technology Innovation Institute (TII). Like Falcon 7B, it predicts text token by token, but its roughly 40 billion parameters make it the larger Falcon entry in the matrix. Do not read 40B as batch size or chip count; it names parameter scale. The practical memory hook is ‘large Falcon from TII.’ TT-Metal lists Falcon 40B with tensor parallelism across eight devices (TP=8) on a Wormhole QuietBox and batch 32. That differs from Falcon 7B’s data-parallel scaling rows: here the larger model itself is partitioned across chips rather than merely replicated for more requests.