# kyegomez/BitNet

Implementation of "BitNet: Scaling 1-bit Transformers for Large Language Models" in pytorch

Repository: https://github.com/kyegomez/BitNet
Canonical: https://ross.abutalabs.com/products/kyegomez-bitnet
Homepage: https://discord.gg/qUtxnK2NMf
Language: Python
License: MIT
License Family: permissive
Topics: artificial-intelligence, deep-neural-networks, deeplearning, gpt4, machine-learning, multimodal, multimodal-deep-learning
Last push: 2026-08-24T00:23:24+00:00

## Health v2 (maintenance only)
Score: 72/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 35, longevity 75
- inputs: {"age_days": 1050, "days_push": 10, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1945, forks 172 (observed 2026-08-28T04:05:57.720396+00:00)

## What it is
A PyTorch implementation of the BitNet architecture from the paper 'BitNet: Scaling 1-bit Transformers for Large Language Models', providing BitLinear layers that replace standard nn.Linear with 1-bit quantized projections. It also includes a full BitNet Transformer and work toward the newer 1.58-bit LLM paper.

## Use cases
- implement 1-bit quantized transformer layers in pytorch
- train large language models with bitlinear layers
- reduce LLM memory footprint with 1-bit weights
- experiment with the BitNet paper architecture
- swap nn.Linear for quantized BitLinear modules
- build a transformer with bit attention and grouped query attention

## When to choose
- you want to reproduce or experiment with BitNet / 1-bit LLM research in PyTorch
- you are training a model from scratch and want 1-bit quantized linear layers
- you need a ready-made BitLinear layer or BitNet transformer implementation

## When to avoid
- you want to quantize an already-trained model - BitLinear requires training or finetuning from scratch
- you need a production-ready, bug-free 1.58-bit implementation - parts are still in progress with known dequantization bugs
- you need inference-optimized 1-bit LLM runtimes rather than a research codebase

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, deep-learning, llm-training, transformers
- domain: deep-learning, large-language-models, machine-learning
- platform: python
- tags: quantization, 1-bit-llm, pytorch, bitlinear, transformer-architecture, research-implementation

## Member repositories
- kyegomez/BitNet (main) score 72

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:57.720396+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:07:23.020229+00:00, confidence not recorded.
  - readme: https://github.com/kyegomez/BitNet (fetched 2026-08-28T04:05:57.720396+00:00, sha a319efe3b71d)
  - homepage: https://discord.gg/qUtxnK2NMf (fetched 2026-08-29T10:46:57.022161+00:00, sha 3c791664bd51)
  - registry_pypi: https://pypi.org/pypi/bitnet/json (fetched 2026-08-29T10:46:57.030885+00:00, sha 6bef50479c38)
- Data as of 2026-08-30T08:39:29.467469+00:00.