Monitor this situation.

Unsubscribe anytime.

[SITUATION] · [QUIET] · [TECHNOLOGY]

2 clusters · 7 sources · 3 days · First seen 11 days ago · Last updated 9 days ago

PrismML release of Bonsai 2 27B AI model

Overview

PrismML, a startup founded by Caltech researchers, has released Bonsai 2 27B, an ultra-lightweight multimodal generative AI model. The model is based on Alibaba’s Qwen3.8 27B and utilizes a “ternary weights” technique—using only +1, 0, and -1 values—combined with FP16 group-wise scaling.

This compression method reduces the model footprint from approximately 56 GB to 5.9 GB, making it more than nine times smaller than the full-precision version while retaining roughly 98% of the original benchmark performance. The model supports a 262K-token context window and multimodal text-and-image inputs, designed to run locally on consumer-grade hardware like Apple M-series chips and Nvidia GeForce GPUs to enhance privacy and reduce cloud dependency.

Following the release, it was noted that PrismML has secured $22.25 million in seed funding from investors including Khosla Ventures, Cerberus Capital, and Caltech. The company intends to scale this compression technology to models containing hundreds of billions of parameters.

Entities

Caltech · PrismML · Babak Hassibi · Qwen3.8 · Nvidia

Timeline

9 days ago

[TECHNOLOGY] 2 sources
PrismML launches Bonsai 2 27B lightweight AI model
PrismML has launched Bonsai 2 27B, an AI model that uses ternary weights to compress large models by up to 10x, enabling high-performance AI to run locally on PCs and smartphones.
11 days ago

[TECHNOLOGY] 5 sources
PrismML releases Bonsai 2 27B for local AI execution
PrismML has launched Ternary Bonsai 2 27B, a compressed LLM that retains 98.2% of its original performance while reducing its footprint to 5.9 GB for local execution on mobile and PC hardware.

Sources

developpez.net · espiganoticias.net · games.yahoo.com.tw · hackernews.com · ithome.com · siliconangle.com · techxeber.az