Model Index
Twenty-two models across three families, hosted on
Hugging Face. All are
Apache 2.0 except Lucy-2, which is research-only
because its training data is. Base models continue text and do not
follow instructions. Instruct models are finetuned for
conversation. GGUF builds are quantized for
llama.cpp.
Boris 2 — preview, latest
Checkpoints from the Boris-2 run, which continues to 200B tokens, each
with a short learning-rate decay applied. Custom architecture: load with
trust_remote_code=True. See Boris.
Boris 1.7
A dense trunk with a 30M-parameter n-gram lookup table spliced into the
residual stream. One experiment, one variable — see
Boris.
Boris 1.3
Boris 1.0 — original release
Lucy — image models
| Model | Type | Params | Released |
|---|
| Lucy-2 | Text to image (research only) | 112.5M | 2026.08.12 |
| Lucy | Unconditional faces | 71.4M | 2026.08.09 |
littlerock
Which one should you use? For conversation at the best quality we
currently offer, take Boris-1.3-125M-Instruct.
For text continuation, take Boris-2-Preview-2, the strongest base
model we have published — a research preview with custom modeling
code and, so far, no instruct or GGUF build. For a plain GPT-2 checkpoint
to finetune, take
Boris-1.3-125M. For CPU-only or embedded
use, take a GGUF build. For scaling research at
the smallest end, take littlerock-1M.
For images, take Lucy or Lucy-2. Boris-1.7 is an
architecture experiment rather than a general-purpose upgrade — take it
if the n-gram table is what interests you, not if you want the best text
quality we ship.
