Gradio names three small-model apps as OpenAI Codex prize winners

Why it matters

The winners show coding agents and local models converging into a practical build stack: small teams can ship private, inspectable AI tools without relying on a large hosted model for every user request.

Gradio named three small-model applications as winners of the $10,000 OpenAI Codex prize in its Build Small hackathon, led by an accessibility tool that uses compact vision, language and speech models to narrate web content.

The Hugging Face-owned framework announced the results in a five-post thread on X on August 3rd. Gradio awarded $5,000 to Tiny Narrator, $3,000 to Blood Test Explainer and $2,000 to Uta Sensei.

Gradio emerged from a problem co-founder Abubakar Abid encountered as a Stanford electrical engineering PhD student. Abid wanted a doctor collaborating on a medical computer-vision project to test a model without installing Python, leading him to build a browser interface for machine-learning demos. He recruited Ali Abdalla, Ali Abid and Dawood Khan to release Gradio in 2019. Hugging Face acquired Gradio in December 2021, after users had built more than 300,000 demos with the framework, according to Abid's announcement at the time.

Three applications built around smaller models

Tiny Narrator, submitted by GitHub user cvpfus, won first place with a screen-reader prototype that combines language, vision and text-to-speech components. Its public repository identifies a 3.97-billion-parameter Nemotron model as its main "reader brain," MiniCPM-V 4.6 for image descriptions and the 82-million-parameter Kokoro model for speech.

The interface builds a reading queue from article headings, paragraphs and images. Users can navigate with keyboard commands, request summaries and review a transcript of the generated narration. The code also includes deterministic fallback descriptions intended to keep the reader usable when a model endpoint is unavailable. Tiny Narrator's repository had one GitHub star and no forks when reviewed on August 3rd, indicating that the award reflects hackathon judging rather than established adoption.

Second-place Blood Test Explainer was built by Roman Dolgopolyi and Dimitris Kalligaridis, researchers associated with the American College of Greece's Deree AI Lab. Users upload a blood-test PDF or image, which the app processes with MiniCPM-V 4.6 before converting extracted markers and reference ranges into structured JSON. A curated knowledge base then supplies plain-language explanations.

The builders say the Blood Test Explainer repository contains no external inference call in the product path, keeping uploaded health data within the Space or local machine running the model. Their own evaluation also exposed the limits of narrow fine-tuning: an early schema-focused training run pushed field-level F1 from 0.655 down to 0.078. A later fine-tune using 100 medical-reasoning examples reached 0.746, according to results published by the builders. The repository warns that the application is educational, can misread noisy scans and should not replace a clinician.

Uta Sensei, built by a developer who posts as @odincode, received third place. The Uta Sensei repository describes a Japanese-language tutor that converts song lyrics into lessons containing furigana, vocabulary, grammar notes and grounded question-and-answer sessions. Users can export the results as lesson files, Markdown or Anki vocabulary decks.

Uta Sensei runs the 3.35-billion-parameter Tiny Aya Global model through llama.cpp, with the public version hosted on a basic CPU instance. The developer tested a fine-tuned adapter during the hackathon, then shipped the stock model after it performed better on a fixed lesson benchmark.

Codex was judged as part of the development process

The Build Small field guide required every participating model to remain below 32 billion parameters and every submission to ship as a Gradio app. The competition advertised a $48,000 cash pool, 29 prize paths, 20,000 Modal credits and two Nvidia RTX 5080 GPUs.

OpenAI's category required Codex-attributed commits in a connected GitHub repository or Hugging Face Space. The rules said judges would rank extensive use of Codex across tasks such as model fine-tuning and agent development above lighter coding assistance. Blood Test Explainer's builders said they used Codex to manage commits and pull requests while changing the model architecture. Uta Sensei's creator credited Codex across implementation, testing, frontend development, evaluation and debugging.

Gradio's final payouts also resolve an inconsistency in the published prize table. The field guide lists $5,000 for first place, $3,000 for second and $1,000 for third, even though it describes the Codex pool as totaling $10,000. Gradio's August 3rd announcement assigned $2,000 to Uta Sensei, bringing the three awards to the stated total.

The winning applications share a common technical choice: each puts a small model close to the user and keeps core functions inspectable through public code. Codex sits one layer earlier in that stack, helping individual developers produce and revise the software around those models under a deadline.