Breaking

PrismML Releases Bonsai 2 27B, a Ternary 27B Model Designed to Run Locally in About 5.9GB

September 26, 2026 at 14:16 EDT

Foundation Models
Open Source
AI Agents

PrismML has released Ternary Bonsai 2 27B, a compressed open-weight model that it says preserves 98.2% of the aggregate benchmark performance of Qwen3.8 27B while shrinking the language-model footprint to 5.9GB. The release puts a 27B-class multimodal model within the memory range of consumer hardware, including NVIDIA GPUs through CUDA and Apple devices through MLX, rather than requiring a cloud-hosted full-precision deployment.

September 17, 2026 · PrismML

A 27B model squeezed into a laptop-class footprint

Bonsai 2 27B uses ternary weights to shrink Qwen3.8 27B from full-precision scale to about 5.9GB while preserving 98.2% of aggregate benchmark performance.

5.9GB

language-model footprint

98.2%

aggregate FP16 retention

262K

token context window

143

tokens per second on RTX 5090

Footprint compression, drawn to scale

The full-precision model is about 54GB; Bonsai 2 27B is about 5.9GB.

5.9GBBonsai 2 27B

54GBQwen3.8 27B FP16

Why developers care

A 27B-class model can run locally for coding agents, private document work, multimodal debugging, and hybrid local-cloud workflows.

Where caution remains

Testers have flagged weaker behavior in long runs, tool-heavy loops, OCR, vision, and fine-grained factual recall.

The deployment shift

Ternary compression→Local 27B inference→Task-specific validation

The central bet is useful capability per gigabyte, not simply a higher raw benchmark score.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.