The world's first commercially viable end-to-end 1-bit LLMs
Caltech origins · Apache 2.0 · April 2026
01 — The problem
10B–700B params
GBs of VRAM required
Phones, robots
Vehicles, secure systems
1-bit quantization · ~1 GB footprint · 10× intelligence density
02 — The family
03 — Architecture
Every layer is strictly 1-bit — no mixed-precision rescue weights. Grouped scaling (g128) per layer.
04 — The metric
−log(avg error rate) / model size (GB)
10×+ improvement
05 — Speed
4–5× more energy efficient than 16-bit equivalents
06 — Benchmarks
Bonsai 8B matches or approaches leading 8B-class models across standard evaluation suites:
07 — What becomes possible
Persistent agents running hours without cloud calls
Sub-100ms latency for physical-world decisions
Full capability in remote or bandwidth-constrained environments
Compliant assistants that never send data off-device
08 — Roadmap
8B, 4B, 1.7B models open-sourced under Apache 2.0
Sub-1-bit regimes, hybrid architectures, multimodal and agent-native capabilities
Cloud-scale efficiency gains, open-source ecosystem expansion
09 — Considerations
Noted limitations: Edge cases in long-context reasoning need further evaluation. Hardware-optimized 1-bit kernels still maturing across full device landscape. Standard benchmarks may not fully reflect real-world application performance — domain-specific testing recommended.