Bonsai

#11 in Lütte & Edge-Modelle

prismml · siet März 2026 · 5× · tolest 30. Juni 2026

24

Momentum

Bonsai is a family of 1-bit large language models from PrismML optimized for extremely efficient inference on edge devices. The flagship Bonsai 8B compresses 8.2 billion parameters into 1.15 GB of memory while achieving benchmark scores comparable to standard 8B models. The family uses a proprietary 1-bit architecture where each weight is stored as a binary value ({−1, +1}) with shared scaling factors, achieving a 14-fold reduction in memory requirements compared to 16-bit models.

Belege (5)