News
2026.08.31: Boris-1.7-D60M-n30M
A 60M-parameter dense decoder-only trunk (10 layers, d=768, GQA, SwiGLU,
RoPE, RMSNorm) with a 30M-parameter n-gram embedding table spliced into the
residual stream after block 2, trained on 2.5B tokens in 22h 37m. One
experiment, one variable: does 30M of lookup buy what 30M of dense weight
would at a fixed token budget? The table is exact rather than hashed
— 281,250 bigram and 187,500 trigram rows plus a shared OOV row, one
searchsorted gather per position, no matmul. Ablating it at
inference costs 0.069 bits/char and 0.7–3.1 points across six
benchmarks. It also ships its own 32,768-entry byte-level BPE trained on the
corpus, so it is not a transformers checkpoint and loads via the
loader in its repository. See the Boris page.
2026.08.12: Lucy-2
A 112.5M-parameter text-to-image diffusion transformer pretrained from scratch at 128×128 — rectified flow, DiT backbone, frozen VAE and CLIP text encoder, roughly 44 GPU-hours across a CC3M pretrain and two finetuning stages. Released for research use only, because CelebA is licensed that way and CelebA face crops are what finally fixed faces. Full write-up, including the mandatory negative prompt, on the Lucy page.
2026.08.09: Lucy
A 71.4M-parameter unconditional DDPM generating 128×128 faces, trained from scratch on CelebA for 60,000 steps in 9.54 hours. The released weights are the EMA weights, and the repository carries the training script, the exact launch command, the complete run log, and a fixed-seed EMA preview grid every 3,000 steps.
2026.08.07: littlerock-1M and littlerock-1M-arithmax
A 1,002,720-parameter Llama-architecture model trained from scratch on 5B tokens (40% FineWeb-Edu, 60% DCLM-baseline) in 1 hour 49 minutes, published with an arithmetic-finetuned sibling that serves as a worked example of benchmark specialization surviving a rigorous contamination audit. Full analysis on the littlerock page.
2026.08.01: Boris-1.3-125M-Instruct and GGUF builds
Instruction-tuned on smol-smoltalk and OASST1 over 5.6 hours, with quantized GGUF builds for llama.cpp published the same day.
2026.07.31: Boris-1.3-125M
Seven continued-pretraining passes on the Boris-125M base checkpoint, adding ~2.66B tokens — about 106% more than the original 2.50B-token pretraining run — and raising Mean-6 from 37.88 to 39.38. Leading with FineWeb-Edu before DCLM avoided the ARC regression seen at 75M.
2026.07.26: Boris-1.3-75M and Instruct
Three continued-pretraining passes bring the 75M model to ~3.95B total training tokens and a Mean-6 of 36.79, with instruct and GGUF variants alongside.
2026.07.23: Boris-250M, and GGUF builds across the board
The largest Boris to date, in base and instruct form, plus quantized GGUF builds for the 75M, 125M, and 250M instruct models.
2026.07.18: Boris-75M and Boris-75M-Instruct
A 77.4M-parameter sibling to the original release.
2026.07.15: NMAI becomes OpenCerebral; first release
New Millennium Artificial Intelligence, a division of the Klondike Software Project, is renamed OpenCerebral. The organization, models, and maintainers are unchanged. The same day, Boris-125M and Boris-125M-Instruct are published as the project's first public models.
References to NMAI and to the old KSP-NMAI repository paths
— which still appear in some model cards — refer to
OpenCerebral.