News / AI & Data
Julia 1: The Open-Source Decision Model with 144.3 Million Parameters Running on CPU Published on 28 September 2026 by Christ-loisele (4 min read)
Supersonic Labs has released Julia 1, an open-source decision model running on CPU and based on mmBERT-small, delivering performance comparable to proprietary solutions like Jev. Its implications for local businesses remain to be explored.
A Lightweight and Accessible Decision Model
Julia 1, developed by Supersonic Labs, stands out as an open decision model with 144.3 million parameters designed to run on standard processors, without requiring a GPU. Unlike generative models from OpenAI or Google, Julia 1 does not generate text but selects an answer from a predefined list of 2 to 20 options, assigning probabilities via a softmax mechanism. According to Supersonic Labs, this approach targets use cases where decisions are structured, such as choosing among multiple responses or binary classification (yes/no).
The model is based on mmBERT-small , a multilingual encoder from JHU CLSP trained on over 1,800 languages, giving it theoretical capability to process content in varied linguistic contexts. Its weights are published under the Apache 2.0 license on Hugging Face, allowing for commercial use, modification, and free redistribution. The total cost of its training amounts to approximately $104 in cloud GPU expenses, according to data published by Supersonic Labs and reported by MarkTechPost .
A Structured Decision Model, 1,400 Times Smaller Than a State-of-the-Art System, Can Compete with Proprietary Solutions on Targeted Tasks.
Illustration: Lawing Tech Mixed Performance Against Benchmarks Like Jev
Julia 1 was evaluated on several benchmarks, yielding mixed results. On the Typed Decisions test, it achieved 73.15% accuracy, slightly outperforming Jev (72.70%), a proprietary decision model cited as a reference by Supersonic Labs. However, on the Banking77 benchmark, Julia 1 only reached 64% accuracy, far behind Jev’s 87%. These gaps highlight its limitations in complex tasks, such as understanding intentions in banking exchanges, as indicated by Veriwire .
In terms of latency, Julia 1 demonstrates remarkable performance: a median of 33.15 ms on an Apple M4 processor, with the ability to drop to 12 ms under optimized conditions (particularly via ONNX Runtime), according to measurements published by Supersonic Labs. This speed makes it a candidate for real-time applications, such as embedded assistants or edge computing systems. The model supports up to 8,192 combined tokens, though benchmarks were conducted with 1,024 tokens.
An Architecture Optimized for Lightweight Computing and Local Deployment
Julia 1 distinguishes itself through compatibility with modest hardware environments. Its FP32 memory footprint is 550.5 MB, but it can be quantized to INT8 to occupy just 36 MB, with a precision loss of less than 1.2%. This compression enables its execution on devices with as little as 4 GB of RAM, such as a Raspberry Pi, according to tests reported by AI Radar . Supersonic Labs emphasizes that this feature opens possibilities for edge computing use cases, where decisions must be made locally to avoid risks associated with transmitting sensitive data.
The model is not designed to generate text, but to choose between predefined options, making it a tool suited for structured decision-making processes. For example, it could be used to validate responses in forms, classify documents, or prioritize alerts. Supersonic Labs clarifies, however, that Julia 1 is not a fine-tuned version of Qwen, another popular model, but an original architecture.
What this changes here
For businesses and government agencies in Benin and West Africa, Julia 1 could represent a cost-effective and local alternative to proprietary decision-making solutions, which are often expensive in terms of infrastructure and licensing. Its execution on CPU and low memory footprint would allow organizations with limited resources to deploy decision-support systems without investing in dedicated servers or GPUs. For instance, banks or insurers could use it to automate tasks such as loan validation or fraud detection, leveraging its compatibility with lightweight environments.
Support for over 1,800 languages via mmBERT-small could also appeal to sectors like education or healthcare, where multilingual tools are needed to serve Francophone, Yoruba, Fon, and other populations. However, fine-tuning adjustments would likely still be necessary to optimize its performance for specific languages or domains, as suggested by Supersonic Labs for Chinese. Finally, its potential for edge computing could attract local industries like food processing or manufacturing, where quick decisions on production lines could be automated without relying on external cloud services.
That said, Julia 1’s uneven performance on benchmarks like Banking77 suggests that its adoption should be accompanied by rigorous evaluations, particularly in contexts where accuracy is critical, such as finance or regulatory compliance.
Sources 
