Laptop with code on desk — Photo by Daniil Komov on Pexels
2 min read

PrismML’s Bonsai 2: a 27B brain that fits in 5.9GB without the usual dumb tax

A full 27B multimodal model that fits in the space of a mediocre photo dump — and PrismML says it barely got dumber. THE GIST PrismML launched Ternary Bonsai 2 27B on 17 September — a ternary compress of Qwen3.8 27B at 5.9GB with 262K context and image+text input. It claims 98.2% benchmark retention versus FP16 (~9× smaller), Apache 2.0 weights, and CUDA plus Apple MLX runtimes. THE TAKE Compression that keeps 98% of the original is the difference between a demo and a deploy

THE HOOK

A full 27B multimodal model that fits in the space of a mediocre photo dump — and PrismML says it barely got dumber.

THE GIST

PrismML launched Ternary Bonsai 2 27B on 17 September — a ternary compress of Qwen3.8 27B at 5.9GB with 262K context and image+text input. It claims 98.2% benchmark retention versus FP16 (~9× smaller), Apache 2.0 weights, and CUDA plus Apple MLX runtimes.

THE TAKE

Compression that keeps 98% of the original is the difference between a demo and a deployment. Ternary weights plus custom kernels shift the fight from “can it run locally” to “will teams trust a 1.76-bit brain for agent work.” If retention holds outside PrismML’s suite, cloud tokens just got optional for a lot of desk-side jobs.

OPINION

Near-lossless ternary at 27B is the story SiliconANGLE and PrismML are selling — verify on your evals before you yank API spend. Start with the official write-up and the GGUF drop.

  • Story type: Important story
  • Impact: Medium
  • Confidence: High (official announcement)
  • Audience: Developers
  • What changed: PrismML released Apache-licensed Ternary Bonsai 2 27B at 5.9GB with claimed 98.2% retention
  • What to watch: Independent local throughput and quality comparisons versus Qwen3.8 FP16 and competing quants
  • Reading time: 2 minutes

Continue reading