Ternary Bonsai 2 is a new family of language models, featuring 2B, 4B, and 9B parameter sizes, built upon the Qwen 3.5 architecture. These models are designed to potentially improve upon the performance of standard Qwen 3.5 models, particularly in areas like memory and response consistency, addressing previous looping behaviors observed in earlier Bonsai models. The Bonsai 2 family requires a specific fork of llama.cpp and offers support for dFlash, MTP, and MLX for enhanced performance. AI
IMPACT Offers potentially improved performance and consistency over base Qwen 3.5 models for local deployments.
RANK_REASON Release of a new family of language models based on an existing architecture. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →