PrismML announced 1-bit and ternary versions of its Bonsai Image 4B diffusion transformer designed for local inference on consumer hardware, detailing the release in a thread on X and a supporting blog post.
https://x.com/PrismML/status/2059339157600969199
The 1-bit Bonsai Image 4B weighs 0.93GB, which PrismML says is 8.3x smaller than its full-precision counterpart. A 1.21GB ternary variant uses weights {-1, 0, +1} for more representational flexibility, landing at 6.4x smaller while targeting better visual quality and prompt fidelity. The models are positioned for laptops and phones rather than datacenter GPUs.
"Compression only matters if the model remains useful," PrismML wrote on X, adding that Bonsai Image 4B remains competitive with much larger models on object composition, aesthetic quality, dense prompt following, and human preference metrics.
Alongside the models, PrismML launched Bonsai Studio for iPhone to run image generation fully on-device with no subscription or API calls. The company says both the 1-bit and ternary Bonsai Image 4B models are released under the Apache 2.0 license, with a browser demo on Hugging Face Spaces and model access via its Hugging Face collection.
