Download one 1.19 GB file, open it on an ordinary laptop, and it answers you โ no graphics card, no account, nothing to rent.
Microsoftโs BitNet stores every brain-weight as one of three values: โ1, 0, +1. A tiny lookup turns those values into words using only addition. Thatโs why the file is smaller than most phone games.
What you actually get
BitNet โ Microsoftโs 1-bit AI engine, MIT-licensed, ~42k stars.
The model you can download today: 2.4B parameters, 1.19 GB โ about 1.5 GB of RAM while it runs.
Measured: 1.37โ5.07ร faster on Apple CPUs, 2.37โ6.17ร on Intel, with 55โ82% less energy than ordinary 16-bit AI.
The win is the map, not the model: three values plus a lookup beat stacking more graphics cards.
See it before you install
Official live demo โ type a question, get an answer.
2.31 MB toy โ 11 million parameters, and you can watch its brain as โ1 / 0 / +1 signs.
0xBitNet chat โ the full 2B model runs inside your browser; nothing leaves your machine.
Speed scoreboard โ real tokens-per-second numbers on a cheap machine.
Put it on your laptop
BitRun โ one install line, prebuilt files, a speed dashboard, and a 79 MB starter model.
Official build โ clone the repo, then:
git clone --recursive https://github.com/microsoft/BitNet.git
cd BitNet
conda create -n bitnet-cpp python=3.10 && conda activate bitnet-cpp
pip install -r requirements.txt
huggingface-cli download microsoft/BitNet-b1.58-2B-4T-gguf --local-dir models/BitNet-b1.58-2B-4T
python setup_env.py -md models/BitNet-b1.58-2B-4T -q i2_s
Fast mode needs AVX2 on x86, or NEON on ARM.
Push it further
FairyFuse โ same three values, a kernel that only adds and subtracts: 29.6ร faster kernel, 32.4 tokens/s on one Intel server chip. Authorsโ measurements, early paper.
Litespark โ one pip install, then load a real 1-bit model; up to 18ร throughput on Apple silicon. Early paper.
T-MAC โ the lookup trick now inside bitnet.cpp: 3.4ร on an M2 Ultra, 5.8ร on a Raspberry Pi 5. Peer-reviewed.
DeepSeek-R1 671B โ a real 671-billion-parameter model packed into 131 GB; works on ~20 GB of RAM, slowly.
LittleBit โ research pack putting Llama2-13B under 0.9 GB. Early research, quality drops.
Bigger than your RAM
SSD offload โ a 25 GB model on a 16 GB laptop with two extra launch flags.
PocketLLM โ stream a 30B model off a USB stick in 6 GB of RAM. New project.
solid.cpp โ a one-line patch that lets a 93 GB model page from disk on a 42 GB machine. New project.
DistAi โ pool three old phones over USB into one 14B machine; keep it on a private network.
ik_llama.cpp โ one -rtr 1 flag gave up to 7ร faster prompt reading on the same CPU.
No machine? Use free compute
Cloudflare Workers AI โ 10,000 free units a day, roughly 49,000 output tokens on a 70B model, no card.
GitHub Models โ your GitHub login is the key; 40+ models from an Action.
Hugging Face thin Space โ host a big-model demo that bills the visitor instead of your free quota.
Kaggle โ 30 h/week of free 32 GB graphics cards ยท GitHub Actions โ free CPUs you can fan out.
Read once before you start
The โ100-billionโ line going around the web is a benchmark setting in the paper, not a download. The model you can actually get is 2.4B / 1.19 GB; that benchmark ran 6.58 tokens/s on a Mac Studio, 1.70 on an Intel laptop (source).
The official source has a one-line activation fix in PR #642 โ applying it restores the intended answer quality (perplexity 17.1 instead of 99.8). Windows builders are tracking PR #644.
For loading the model yourself, use bitnet.cpp or the browser demos โ app support is still arriving (request).
FairyFuse and Litespark are early papers โ the numbers are the authorsโ own measurements.
Free compute tiers are for your own experiments; keep them personal.

!