OneBit
We train a new species of AI model, small by design rather than shrunk afterwards. It holds the most intelligence per byte, on the chips you already own, not in a data center.
On any device with a processor: phones, laptops, servers, cameras, drones, robots, satellites.
Runs on:
Any CPUAny CPUThe next Moore's law
For fifty years, more intelligence meant more transistors. The next curve is how much intelligence fits in a byte.
10×
A tenth of the space, the same intelligence
The mostintelligenceper byte.
OneBit
We train a new species of AI model, small by design rather than shrunk afterwards. It holds the most intelligence per byte, on the chips you already own, not in a data center.
On any device with a processor: phones, laptops, servers, cameras, drones, robots, satellites.
Runs on:
Any CPUAny CPUThe next Moore's law
For fifty years, more intelligence meant more transistors. The next curve is how much intelligence fits in a byte.
10×
A tenth of the space, the same intelligence
The mostintelligenceper byte.
The idea
From 65,536 shades to three
Shades per cell
65,536
Space it takes
100%
A language model is a web of billions of connections. What it knows is the shape of that web: which connections exist, which way each one pushes, and roughly how hard. Today every connection is stored as a high-precision number, and the precision is the expensive part. It is why the model fills a data center, why it needs a GPU, and why you reach it through a wire.
The precision is almost entirely unused. The intelligence is in the pattern, not in the decimals. A connection needs to say push, pull or stay out of it, and very little more. Written that way, the same web takes about a tenth of the space, and the arithmetic collapses to addition, which every processor made in the last decade does quickly.
That is what OneBit builds: models trained from the first step to live in that form, rather than large models squeezed into it afterwards. The intelligence is the same. The places it can live are not.
Where it runs
On the machine itself. No data center.
What OneBit does to a model
| Original, FP16 | OneBit, ternary | |
|---|---|---|
| Weights on disk | 8.90 GB | 2.66 GB |
| Bits per weight | 16 | ~1.94 |
| Capability kept | 100% | 87%, target 100% |
| Hardware | GPU | any CPU |
| Inference | in a data center | on the device |
| Cost per token | metered | none |
| Data | cloud | local |
Qwen3-4B in FP16 against Cloe 1.2, its ternary conversion. Numbers from the paper.
Where the answer is made
As models are stored today
8.9 GB
Cloe 1.2, converted after training
2.66 GB
Trained small from the start
About 0.9 GB
Intelligenceat theedge.
The four reasons intelligence can leave the data center.
What a model knows is a map
A model is billions of connections. Each one strengthens a signal, weakens it, or leaves it alone. What the model knows is that map: which connections exist and which way they push. The exact strength is stored to a precision the map never uses.
- Billions of connections
- Strengthen, weaken, or leave alone
- The map is the knowledge
What a model knows is a map
A model is billions of connections. Each one strengthens a signal, weakens it, or leaves it alone. What the model knows is that map: which connections exist and which way they push. The exact strength is stored to a precision the map never uses.
- Billions of connections
- Strengthen, weaken, or leave alone
- The map is the knowledge
Most of the detail is waste
Keep the map and drop the precision, and the model keeps working. The same model takes about a tenth of the space, and its arithmetic turns from multiplication into addition.
- A tenth of the space
- Addition instead of multiplication
- Nothing lost that matters
Most of the detail is waste
Keep the map and drop the precision, and the model keeps working. The same model takes about a tenth of the space, and its arithmetic turns from multiplication into addition.
- A tenth of the space
- Addition instead of multiplication
- Nothing lost that matters
Ordinary chips are enough
Addition is what ordinary processors are good at. A model in this form fits in the memory of a laptop, a server or a phone, and runs on the chip that is already there. No data center, no internet. A GPU is optional, for even faster inference.
- Any CPU
- Works offline
- GPU optional
Ordinary chips are enough
Addition is what ordinary processors are good at. A model in this form fits in the memory of a laptop, a server or a phone, and runs on the chip that is already there. No data center, no internet. A GPU is optional, for even faster inference.
- Any CPU
- Works offline
- GPU optional
Anything with a chip can think
When intelligence is this small it stops being a service you rent and becomes a part you build in. Cameras, drones, robots, satellites, factory machines: anything with a chip can carry its own.
- Built in, not rented
- Cameras, drones, robots, satellites
- Every chip, eventually
Anything with a chip can think
When intelligence is this small it stops being a service you rent and becomes a part you build in. Cameras, drones, robots, satellites, factory machines: anything with a chip can carry its own.
- Built in, not rented
- Cameras, drones, robots, satellites
- Every chip, eventually
Models
Weights on Hugging Face, code on GitHub, methodology in the papers. Everything we ship, you can run.
Open weights
Cloe 1.2Cloe 1.2Cloe 1.1S1.0
01 – 03
Size on disk · was 8.90 GB
2.66 GB
Size on disk · was 3,037 MB
635 MB
Size on disk
60 MB
Weights on Hugging Face, code on GitHub, methodology in the papers. Everything we ship, you can run.
Scaling trajectory
Models converted after training keep most of what they knew. A model trained small from the first step keeps all of it, so every byte it sheds is pure gain. That is the line we are on, and it does not stop at ten.
↑ Intelligence per byte, relative to the original
→ Times smaller than the original
- The original model · 1×
- Cloe 1.2 · converted · 3.4× smaller · 87% kept
- Cloe 1.1 · converted · 4.8× smaller · 77% kept
- OneBit-T 3B · trained small · 4.3× smaller · 103% kept
- Projected · trained small · 10× smaller
Intelligence per byte = capability kept × size saved. Dashed parts are extrapolated.
Papers
01 – 05
Post-Training Ternarization of Qwen3 Language Models
Rotation, ternarization and error compensation on Qwen3-4B, end to end: capability, effective bits per weight, storage and inference, measured.
Malik, Devan, Mehra
Technical report · August 2026
01 – 05
Team
Compressing intelligencefrom first principles.
The people
A research lab in Delhi, with people in Zurich and San Francisco: physicists and engineers working out how much intelligence a byte can hold.
The people
Team
Compressing intelligencefrom first principles.
A research lab in Delhi, with people in Zurich and San Francisco: physicists and engineers working out how much intelligence a byte can hold.
Get early access.