Supersonic Labs
Supersonic Labs is a small independent AI research lab in Brazil focused on making practical, understandable intelligence work on hardware people already own. It builds compact models and open systems for concrete tasks such as classification, routing, search, and extraction.
At a Glance
- Developers integrating compact decision models into existing software
- Teams needing local or edge AI on CPUs, tablets, or browser hardware
- Organizations with classification, routing, prioritization, or extraction workflows
- Researchers and open-source AI practitioners evaluating reproducible decision models
AI Tools by Supersonic Labs
(1)
Julia 1
Multilingual CPU Decision Model
Discussions
No discussions yet
Be the first to start a discussion about Supersonic Labs
Latest News
Supersonic Labs releases Julia 1, a 144.3M-parameter open decision model that runs on a CPU
Supersonic Labs announces Julia 1 as its first classification model designed to run on almost anything
Julia 1 CPU evaluation reports 1,451/2,000 typed decisions correct and a 60/100 Banking77 result
Julia 1 evaluation reports results across typed decisions, AG News, DAIR Emotion, Banking77, and MASSIVE
Products & Services
A 144.3-million-parameter multilingual decision model built on mmBERT-small. It takes context, a question, and 2–20 possible answers, then selects an answer and returns scores/probabilities in the supplied order. It is designed to run locally on CPUs and is distributed with Python support, weights, metrics, and provenance under Apache 2.0.
A browser-oriented ONNX export of Julia 1 for WebGPU inference. It uses ONNX Runtime WebGPU and a bundled Rust WebAssembly tokenizer, keeps computation in the browser, and exposes the same choice, score, and yes/no decision interface.
Market Position
Supersonic Labs positions Julia 1 as an open-weight, CPU-capable decision model rather than a generative chatbot. Relative to hosted TypeSafe Jev, it emphasizes local deployment, Apache 2.0 access, and a low planned token price; relative to broader open models such as GLiNER2.5 Multi, it offers a narrower, explicit choice/score/yes-no interface. Its trade-off is that it does not generate text or supply missing knowledge, and its hierarchical routing showed a substantial weakness on the 72-label Banking77 pilot.
Leadership
Founders
João Davi
CEO and founder of Supersonic Labs; his public profile describes his focus as bringing capable AI models to hardware people already own.
Júlia Klee
Co-founder, Head of Research, and Model Architect at Supersonic Labs; she leads research and model architecture for the lab.
Executive Team
João Davi
CEO and Founder
Publicly describes his work as bringing powerful AI models to hardware people already own.
Júlia Klee
Co-founder, Head of Research and Model Architect
Leads research and model architecture at Supersonic Labs.
Founding Story
The lab was started around the belief that useful intelligence should be close at hand: able to work where people live and build, respect the limits of their machines, and earn trust through measurable usefulness. Its manifesto emphasizes starting with real tasks, choosing the smallest system that can do the work well, and sharing reproducible work so others can build on it.
Business Model
Revenue Model
The lab currently distributes open model weights and code, while developing a hosted API for Julia 1. The planned API is usage-priced by input tokens; the model page says access is not yet open.
Pricing Tiers
Planned launch price; the hosted API was announced as not yet open.
Julia 1 and the WebGPU/ONNX build can be run locally; infrastructure and hardware costs are borne by the user.
Target Markets
- Developers integrating compact decision models into existing software
- Teams needing local or edge AI on CPUs, tablets, or browser hardware
- Organizations with classification, routing, prioritization, or extraction workflows
- Researchers and open-source AI practitioners evaluating reproducible decision models
- Text classification
- Support and request routing
- Ordered prioritization or severity scoring
- Binary/yes-or-no decisions
- On-device or edge inference where CPU execution, latency, memory, or privacy matter
- Browser-based decision inference through WebGPU