PrismML Releases Ternary Bonsai 2 27B: 5.9 GB Model Retaining 98.2% Performance

Ternary Bonsai 2 represents a significant breakthrough in model compression for edge deployment. By fitting a capable 27B parameter model into just 5.9 GB while retaining 98.2% of the original Qwen3.8 27B model's performance, this work pushes the boundaries of what's practical for on-device inference. The extreme compression is achieved through ternary quantization techniques that reduce each parameter to ternary values while maintaining remarkable capability preservation.

For local deployment, this model opens new possibilities on resource-constrained devices. A 5.9 GB model fits on modern smartphones with room to spare, yet maintains strong performance on reasoning and language tasks. The Apache 2.0 license ensures commercial viability. This represents the kind of practical breakthrough that moves local LLMs from research curiosity to mainstream consumer feature capability.

Read the full article on Google News.

Source: Google News · Relevance: 9/10