The presented Ternary Bonsai 2 9B model in GGUF format, occupying only 2.4GB, has demonstrated the ability to generate animation code based on an image and a text prompt. During testing, the model was given an image of a doner kebab and a corresponding prompt, after which it generated code to create a rotating animation of the dish.

Fahad Mirza, the author of the review, notes that the Bonsai family of models shows variable results: sometimes they work excellently, and sometimes they fail to complete the assigned tasks. He expressed interest in the performance of the new Bonsai 2 9B model.

About Ternary Bonsai and Prism ML Models

Ternary Bonsai 2 9B is a 9-billion parameter model in GGUF format. It is based on the Qwen 3.5 9B model and has the capability to process both images and text while maintaining a context window of 262K. The original Qwen model is approximately 17.9 GB in size, whereas Ternary Bonsai 2 9B is only 2.4 GB, which is about seven times smaller. The model weights are distributed under the Apache 2.0 license, allowing for free usage.

The model compression process involves converting each weight from a 16-bit number (e.g., 0.82 or -0.07) into one of three values: -1, 0, or +1. To maintain the model's size, additional information is used – a numerical value for each group of 128 weights. Reducing the number of bits per weight leads to a significant reduction in file size.

Performance Testing

The model was run using llama.cpp. Video memory (VRAM) consumption at a full 262K context window was recorded. Reducing context length by changing the C parameter allows for lower VRAM consumption.

Nadsat Slang Knowledge Test

As part of one test, the model was tasked with retelling a real historical event in the Nadsat slang from "A Clockwork Orange." The goal was to check if the model retained knowledge of the original slang or merely generated poetic-sounding text.

The result was assessed as 50/50. The model used Nadsat words, but, according to the author, in most cases with incorrect meanings. Only the words "drug," "malchik," and "moloko" were deemed correct. All 11 words used were taken from the original prompt. The author notes that the model "sprinkles" words but does not "speak" Nadsat.

Multilingualism Test

The next test aimed to verify the model's multilingual capabilities. The model was presented with a restaurant scenario and asked to respond in 28 languages.

Major languages such as Spanish, Portuguese, and Mandarin were handled well. However, minor errors were noted in Korean and Indonesian. Medium-resource languages, including Hindi, Urdu, Greek, and some African languages, were presented in a distorted form with interspersed Russian and Chinese words. Rare languages, including European, North American, and others, were mostly processed satisfactorily, but some rare languages "collapsed" into a repetitive, nonsensical string of characters. The author concludes that the quality of responses decreases as the language becomes more complex.

Scientific Reasoning Test

For this test, the model was presented with an infographic of a wind tunnel and asked to analyze it as a senior aerodynamicist. The task involved checking numerical data, physical principles, and mathematical calculations.

Analysis of the model's reasoning showed that it handled the task at a high level, without missing important details. The model correctly identified contradictory Reynolds numbers, detected distorted text and nonsensical formulas, and acknowledged the report as unreliable. However, the level of detail was less precise: the model misread small numbers and invented a stall angle of less than 90 degrees, which was not present in the image. Checking physical principles also revealed an arithmetic error.

Overall, the 2.4 GB model is capable of reasoning about complex diagrams and drawing reasonable conclusions, but its numerical data should not be taken as absolutely accurate. It is noted that hallucinations are present, but their level has decreased compared to previous models in the family.

Situational Test

As part of the final test, the model was presented with an AI-generated image: a wealthy bachelor with a large sauce stain on his shirt, caught between two rival future mothers-in-law—an Italian nonna with lasagna and a baby photo, and a Pakistani aunt with fried chicken and a family album. The model was asked to describe the image, write a funny sketch with cultural references, Italian and Urdu phrases, and an unexpected ending.

The visual description was excellent: the model correctly described all characters, dishes, and the stain. The Italian phrases were adequate, including one humorous line. However, the plotline was weak: the Urdu phrases were inaccurate and repetitive. Daughters and marriage were not mentioned, the man suddenly changed his blue shirt for a suit, and the twist with the baby photo made no sense.

The model demonstrated good image reading skills and satisfactory Italian language proficiency but showed weakness in plot construction and creativity.

In conclusion, the author notes a significant performance loss compared to larger models but acknowledges Prism ML's efforts in creating this model.