Supersonic Labs, an artificial intelligence laboratory based in Brazil, has introduced Julia 1, a novel decision-making model. Unlike traditional chatbots, Julia 1 is designed to receive context, a query, and a selection of two to twenty possible answers, subsequently identifying the most probable option and assigning a probability to each. The model, comprising 144.3 million parameters, is notable for its capability to operate on a conventional CPU.
Deployment and Accessibility
Julia 1 is designed for broad deployability. Its weights are publicly available on Hugging Face under an Apache 2.0 license, allowing for local execution using Python 3.11+ on a CPU or a BF16-capable GPU. Additionally, an ONNX build enables in-browser functionality via WebGPU. While a hosted API has been announced, it is not yet live for public access.
Core Functionality
The model's API supports three distinct decision types:
- Choice: This function enables the selection of a single label from a described set of 2 to 20 options, suitable for classification or routing tasks.
- Score: Users can obtain an expected index on an ordered rubric, such as low, medium, or high.
- Noul: This feature provides the probability that a given yes-or-no statement is accurate.
Results are returned in the user-specified option order, complete with full softmax probabilities. Input caller IDs, such as for billing purposes, are preserved. It is important to note that Julia 1 does not generate text.
Architectural Foundation and Development
Julia 1's architecture is rooted in JHU CLSP's mmBERT-small, a multilingual ModernBERT encoder with 140 million parameters, trained across over 1,800 languages. Supersonic Labs retained the encoder and tokenizer, incorporating a specialized decision head, and subsequently trained the model using decision-formatted examples. The lab clarifies that Julia 1 is not a fine-tuned Qwen model. The runtime supports up to 8,192 combined tokens, though published benchmarks utilized a 1,024-token limit. The total cloud GPU expenditure for training and experimental phases amounted to approximately R$540 (US$104.08). The FP32 weights occupy 550.5 MiB of storage. Details regarding the private training pipeline have not been disclosed. Development of Julia 2, which will feature the lab's proprietary foundational architecture, is currently underway.
Performance Benchmarks
Evaluations conducted on September 24, 2026, using H200 BF16 with strict encoding, compared Julia 1 against TypeSafe's Jev, utilizing Jev's established benchmark protocol reference values. Key results include:
- Typed Decisions: Julia 1 achieved 73.15% (1,463/2,000) accuracy, surpassing Jev's reference of 72.70%.
- AG News (4 labels): A score of 94/100 was recorded, compared to Jev's 91% reference.
- DAIR Emotion (6 labels): Julia 1 demonstrated 86/100 accuracy, significantly outperforming Jev's 48% reference.
- Banking77 (72 labels): The model scored 64/100, notably trailing Jev's 87% reference. This category represented a clear underperformance.
- MASSIVE (18 scenarios): An overall macro accuracy of 71.50% was observed across 52 locales, with specific high points of 86.25% in pt-PT and 86.75% in en-US.
The classification pilots were based on 100 examples each. A subsequent CPU run on September 25 replicated most figures, including 72.55% on Typed Decisions and 60/100 on Banking77 with three abstentions.
On-Device Latency
Supersonic Labs published specific latency measurements across various devices:
- An Apple M4 recorded a median latency of 33.15 ms per decision.
- A Samsung SM-X510 tablet, utilizing ONNX Runtime, yielded a median of 203 ms with a peak RSS of 393.1 MB.
- On an Intel Core i5-1235U, AG News decisions averaged 107.83 ms.
- Banking77 decisions on the same i5 processor took 3,713.54 ms, primarily due to the initial narrowing process for its 72 labels.
While claims on social media suggest Julia 1 classifies five times faster than Jev on an i5 laptop, it is noted that the Jev pilot measured its performance as a hosted service called from France, meaning the latency comparisons are not directly equivalent.
Competitive Overview
Julia 1 distinguishes itself among competitors such as TypeSafe Jev and GLiNER2.5 Multi, particularly in its open-source license and CPU compatibility. With 144.3 million parameters, it is smaller than GLiNER2.5 Multi's 287 million but larger than Jev's undisclosed parameter count. Julia 1 supports choice, score, and yes/no decision types, offering 2 to 20 options per call, and notably runs locally on a CPU, a feature absent in Jev. Its planned API pricing is set at $0.025 per million tokens, compared to Jev's $0.042.
Key Limitations
Julia 1's capabilities are limited to comparing provided answers; it cannot perform tasks requiring missing facts, algebraic calculations, or multi-step computations. The Router mechanism may inadvertently omit the correct label during the narrowing process. It is not a direct replacement for a Transformers pipeline, and no Hugging Face inference provider currently hosts it. Supersonic Labs advises users to conduct their own evaluations with specific questions and to maintain human oversight for critical decisions.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
