Supersonic Labs has launched Julia 1, an open-source decision model featuring 144.3 million parameters. The model is built upon the mmBERT-small encoder from JHU CLSP, which supports over 1,800 languages. Unlike generative models, Julia 1 is designed specifically for classification tasks and does not produce text output.
The model is available on Hugging Face under the Apache 2.0 license. It supports local execution on CPUs or BF16-capable GPUs using Python 3.11 or higher. Additionally, an ONNX build allows the model to run in web browsers via WebGPU. The developers have also announced plans for a hosted API.
Training and experimentation costs for the model totaled approximately $104.08 in cloud GPU resources. The model weights in FP32 format occupy 550.5 MiB of storage. The runtime environment supports a context window of up to 8,192 tokens, though initial benchmarks were conducted with a 1,024-token limit.
Performance tests on an Apple M4 chip showed a median latency of 33.15 ms per decision. On an Intel Core i5-1235U processor, the model processed AG News tasks with a median time of 107.83 ms. The developers are currently working on Julia 2, which will utilize a proprietary foundation architecture.
