quantization

3 posts

PrismML's Ternary Bonsai 2 27B fits a 27B multimodal model into 5.9GB, and its footprint and energy numbers matter more than parameter count.

Benchmarks show Qwen3.8 27B keeps full quality at 4-bit but collapses to random-guess scores at 1-bit, with guidance on which quant to run locally.

Cerebras quantizes model weights only at storage and runs no pruning on its Shared Inference tier, and promises stable model IDs you can pin to.