OneBit AI develops software that compresses large language models into tiny, ternary representations that run entirely on edge devices like phones, cameras, and microcontrollers—with no cloud connection. Its Cloe 1.2 model shrinks an 8.9 GB Qwen3-4B to 2.66 GB while preserving 87% of capability, and its S1.0 model packs 70 million parameters into just 60 MB. The company targets manufacturers, consumer electronics makers, and IoT developers needing on-device intelligence without GPUs or data centers.
Funding
Funding not disclosed
Founders
Product
Problem
Modern language models are stored as high-precision numbers, requiring gigabytes of memory, powerful GPUs, and cloud data centers to run. This makes them inaccessible for consumer devices, industrial sensors, and microcontrollers where connectivity is limited, power is scarce, and data privacy is critical.
Solution
OneBit AI converts large language models into ternary form, where each weight encodes only push, pull, or stay out of it, cutting model size to roughly one-tenth of the original while keeping most capability. The resulting models run on standard CPUs—inside phones, cameras, drones, robots, and microcontrollers—with no GPU or cloud dependency. Live transcription, defect detection, meter reading, and wake-word recognition all execute locally, with only results transmitted. The company offers models converted post-training (Cloe series) as well as models trained from scratch in ternary space (S1.0), with a projected path to 10× size reduction while maintaining full capability. All benchmarks are published with the exact hardware and method used.
Target Audience
Primary customers are manufacturers building smart phones, TVs, cars, cameras, drones, industrial inspection systems, wearables, toys, and edge IoT devices that need embedded AI without cloud reliance.
Features
- Ternary-weight compression reducing models to ~1.94 bits per weight (Cloe 1.2) and 1.58 bits per weight (Cloe 1.1)
- Post-training ternarization with rotation and error compensation, requiring no retraining for converted models
- Quantization-aware training for models built ternary from the first step, enabling 60 MB models with 70 million parameters
- On-device inference on any CPU, eliminating GPU requirements and cloud data transfer, with models as small as 5–11 MB for speech
- Up to +24% faster decode throughput compared to latent baselines, with all-day battery operation on wearables and multi-year battery life on sensors
- Works entirely offline, with data never leaving the device; only result metadata is transmitted