The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits

Pure 1-bit weights quantize poorly because the +1/-1 grid cannot express that some weights should be near zero.

BitNet b1.58: every weight is TERNARY {-1, 0, +1}, keeping a single bit per sign plus a single shared exponent/bits-per-group field - the '1.58' is that structural overhead. Absorption (folding the RMSNorm scale into the following linear layers) removes the normalization at inference so the ternary matmul is the whole cost. This is again an architecture trained from scratch, not a compression of an existing model.

Paper's own claim (abstract): b1.58 'matches the full-precision (FP16 or BF16) Transformer LLM with the same model size and training tokens in terms of both perplexity and end-task performance', while being significantly more cost-effective in latency, memory, throughput and energy. COMPRESSION RATIO: 1.58 bits/weight (~10x vs FP16, and the storage overhead is in the 0.58). QUALITY COST: the headline claim is MATCH to FP16 - but only when size and token count are equal, which is a training-cost condition, not an inference-cost one. No end-to-end serving throughput number is given in the abstract.

SLUG RENAMED 2026-10-04 (parent agent): was 'bitnet-b158', colliding with the quantization record of the same id, which made every bare-slug cross-reference ambiguous. Inbound refs (['scheme']) all mean the quantization record, so none needed repointing; papers are cited via `sources`, never by bare slug. arXiv ID 2402.17764 verified via /abs/. Venue: Semantic Scholar returns arXiv.org only, arXiv comment empty - arxiv-preprint, year = v1 year 2024. Frequently mis-cited as '1-bit'; the ternary -1/0/+1 grid and the shared exponent field are what the 1.58 accounts for, and the repo's quantization/bitnet-b158 record carries the format detail. arXiv ID and title verified by fetching https://arxiv.org/abs/2402.17764 and reading the citation_title meta tag on 2026-10-03. Venue and year from the Semantic Scholar graph API batch endpoint (paper/arXiv:2402.17764) unless the record's notes say otherwise. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.