翻訳待ち：The Sequence Knowledge- Issue 924: The Distilled Models You Need to Know About

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

In July 2026, PrismML released Bonsai 27B, a model that makes the old relationship between parameter count and hardware look slightly absurd. A conventional 27-billion-parameter model in 16-bit precision needs roughly 54 gigabytes just for its weights. Bonsai ships in a ternary version around 5.9 gigabytes and a binary version around 3.9 gigabytes. The latter is designed to fit inside the memory budget of a high-end phone. It is multimodal, supports long context, and preserves much of the reasoning and tool-use behavior of the full-precision Qwen3.6-27B model from which it was derived. This is a useful place to begin an essay about distillation because Bonsai is not, strictly speaking, a textbook distillation model. PrismML’s public materials emphasize end-to-end low-bit training and quantization rather than a classical teacher-student loss. Bonsai is better understood as a boundary object: a glimpse of the point where distillation, pruning, quantization-aware training, and systems engineering collapse into one manufacturing stack. The interesting unit in modern AI is no longer the individual checkpoint. It is the lineage. A frontier model discovers a capability. A smaller model learns its probability landscape. Another absorbs its reasoning traces. A pruned descendant inherits its architecture. A low-bit version packages the result for a device. The model family is becoming a family tree. What actually counts as distillation? Read more