Latents, long context, open data

Adds product-quantized next-concept prediction to autoregressive decoding, matching OLMo-3-7B loss with 51.3% tokens.

Unlike larger post-trained specialists, it trains an open 7B from scratch to 256K context, ranking first across 14 reasoning benchmarks and reaching 4.2x faster 16K pretraining time-to-loss.

It learns one next-state latent from video and language, then probes frozen backbones for text, images, and robot actions.

Unlike single-method decoding repos, it unifies data prep, multi-GPU training, and evaluation across DSpark, DFlash, Eagle3, and four target-model setups.

Unlike recent typed-decision repos, it packs many question types into one calibrated, decode-free pass with 2.0× speedup and 4e-6 output parity.

Unlike single-modality embedders, it unifies five input types in one space and retains 98.7% image/video performance at 256 dimensions.

Unlike paper-text corpora, it combines 3.15M arXiv papers with source files, resolved TeX, and 5.03M-version history in one auditable 16.08 TB snapshot.

Unlike broad corpora, it tests evidence-grounded QA over documents up to 2,935 pages with page-level evidence, cross-document questions, abstention, and only 38.36% top accuracy.

It offers 11 million CC0 full-text Turkish court decisions spanning 1962–2026 across five court systems.