Artificial intelligence laboratory PrismML has launched a new language model that is small enough to run on ordinary PCs and powerful smartphones. The company believes such systems don’t have to be large.
Image source: BoliviaInteligente / unsplash.com
New features in line language models PrismML Bonsai 2 27B Is a 27B version of the popular open system Alibaba Qwen3.8, but is quite compact in size – only 5.9 GB. Therefore, it can be easily installed on PC and even run on flagship smartphones. The model volume is 9-10 times smaller than the original.
The compression technology used by PrismML is unique, and the developers claim that the company creates models with almost no performance loss compared to the original model. The new Bonsai 2 showed 98% of the cumulative benchmark results shown by Qwen. This number is higher than the first Qwen released in March, which offered 95% of the original version. If the original model was downloaded 11 million times, the compact PrismML model was downloaded 2.6 million times. Therefore, the compression results displayed at startup will improve from version to version.
Whether 100% compliance can be achieved is an open question and of little practical significance. The original model is not very accurate and the benchmarks do not perfectly reflect real-world tasks, and a 2% decrease in efficiency can have a significant impact on real-world performance. Software connectivity remains an important factor.
PrismML explains that impressive results can be achieved by reducing the size of the weights that make up the model. Typically requiring 16 bits each – Enable uses ternary weights, that is, it reduces this volume to three values: “+1”, “0” and “-1”. Since much less data needs to be stored per weight, the model takes up significantly less space. The startup plans to apply this compression method to even larger models – with hundreds of billions of parameters; the developers believe it will be easier to retain “intelligence” in them, since as the model size increases, the compression possibilities expand without losing functionality. It may be easier to achieve 100% compliance with them.
If you find an error, select it with your mouse and press CTRL+ENTER.
