skai-research/lca-3x-base
0.4B • Updated • 8
LCA-MBP: Dynamic Multi-Byte Prediction
374M byte-level models from Dynamic Multi-Byte Prediction With Hierarchical Language Models — arxiv.org/abs/2608.15454
Note LCA-MBP, the paper's method: multi-byte prediction head with Latent Causal Attention, ~3.3 bytes/patch.
skai-research/lca-3x-sft
0.4B • Updated • 9 • 1Note LCA-MBP after Tulu SFT. The instruction-following model behind the speculative decoding results.
skai-research/fxt-3x-base
0.4B • Updated • 10Note Hierarchical byte baseline at the same segmentation rate, without the MBP head.
skai-research/fxt-3x-sft
0.4B • Updated • 12Note The no-MBP baseline after Tulu SFT.
skai-research/bytes-base
0.4B • Updated • 10Note Flat 22-layer byte transformer: no segmentation, one forward pass per byte.
skai-research/bytes-sft
0.4B • Updated • 12Note The flat byte baseline after Tulu SFT.