Small models, big builds: OpenBMB × Hugging Face Build Small Hackathon winners announced
946 runnable AI applications, 817 developers, 263 projects built with MiniCPM. From a local travel agency CRM and a plant health assistant to vibe-based route planning, AR bedtime stories, and a fantasy world that keeps growing—when models get small enough, AI ideas find more room to land.
As the industry keeps chasing larger parameter counts, longer context, and more expensive compute, a developer experiment built around small models offers a different answer.
The Build Small Hackathon, initiated by Hugging Face and Gradio with OpenBMB participating as an Anchor Sponsor, has officially come to a close.
The hackathon did not ask developers to build an all-purpose AI. It set a clear boundary:
Use only small models with 32B total parameters or fewer, build the application as a Gradio App, deploy it to Hugging Face Spaces, and submit a product people can actually try.
Within that boundary, developers had to answer a more important question:
When a model is small enough to be understood, fine-tuned, and deployed by an individual, what real problems can we actually solve with it?
946 projects that turned small models into real products
This edition of the Build Small Hackathon produced 946 runnable AI applications and drew 817 developers. Hugging Face grouped the winning projects into main track awards, sponsor awards, special awards, and community awards.
MiniCPM was the second most popular model family in the event, adopted by more than 27.8% of projects (263)—roughly one in every four.
More notable still, 74 teams trained and released their own fine-tuned models.
Developers did not simply wire a general-purpose model into an interface. They built data, fine-tuned for a domain, distilled models, quantized to GGUF, deployed locally, and optimized inference pipelines around specific tasks.
From an idea to a Hugging Face Space people can open, interact with, and experience, hundreds of developers completed their own small-model adventure in just two weekends.
Two tracks: solve real problems, or create new worlds
The Build Small Hackathon had two main tracks: Backyard AI, focused on concrete problems in real life, and Thousand Token Wood, which encouraged developers to create imaginative AI-native experiences.
OpenBMB offered first, second, and third prizes in each track, with a total prize pool of USD 10,000. In the end, six projects that stood out for product value, creative expression, and depth of MiniCPM use received OpenBMB Awards.
Backyard AI brings small models into real life: 🥇 First prize | Viva Mais AI extracts structured business data from WhatsApp conversations
This article includes video content. Please view it in the original WeChat post.
Viva Mais AI is a localized WhatsApp Travel CRM built for small travel agencies.
It reads flight tickets, orders, payment receipts, customer ID documents, and voice messages from chat history, then turns the scattered information into structured customer profiles, helping agencies track orders, reconcile payments, and answer customer questions.
The project uses MiniCPM-V 4.6 for image and document understanding and MiniCPM5-1B for information extraction and Q&A, and supports local processing of sensitive customer data.
➤ Project link
🔗 https://huggingface.co/spaces/build-small-hackathon/vivamais
🥈 Second prize | Aranya: A WildKeeper’s Adventure turns plant identification into an audible jungle expedition
Aranya is an AI plant assistant that combines plant identification, health analysis, care advice, and voice narration.
The team fine-tuned two specialized models based on MiniCPM-V 4.6, one for plant species identification and one for plant health analysis, and built local inference and voice interaction with llama.cpp and Pocket TTS.
It does not stop at the conventional "take a photo to identify" tool. Through designs such as Discovery, Rescue, and a plant journal, it turns plant knowledge into a jungle exploration that accumulates over time.
➤ Project link
🔗 https://huggingface.co/spaces/build-small-hackathon/aranya_a_wildkeepers_adventure
🥉 Third prize | WanderLust takes you not on the fastest route, but on one you will actually enjoy
WanderLust is a local-first "vibe route" planning assistant.
Users can enter not only a start and destination but also the kind of route experience they want, such as "past bookstores and quiet side streets" or "good for a Sunday morning walk."
The project uses MiniCPM5-1B to understand user preferences, then works with OpenStreetMap and conventional route planning algorithms to generate exploration routes that better match personal interests within a limited extra travel time.
➤ Project link
🔗 https://huggingface.co/spaces/build-small-hackathon/WanderLust
Thousand Token Wood, where small models start creating worlds: 🥇 First prize | Grandpa’s Bedtime Stories turns an ordinary photo into an AR story world you can enter and talk to
Grandpa’s Bedtime Stories is a WebXR-based interactive AR storytelling app.
After uploading an ordinary photo, users can enter a 3D Gaussian Splat scene generated from it, select a person or object in the scene, and ask "Grandpa" questions by voice.
The project uses MiniCPM-o 4.5 to handle visual understanding, speech understanding, story generation, and spoken answers at once, turning a still photo into a story world you can enter, question, and tell stories in together.
➤ Project link
🔗 https://huggingface.co/spaces/build-small-hackathon/grandpas-bedtime-stories
🥈 Second prize | Aether Garden, a magic book that keeps writing itself
Aether Garden is an AI fantasy world that keeps growing.
Users can create characters and leave them in the world for good. Even after the user leaves, the characters keep meeting, forming relationships, and living through new events.
The project uses MiniCPM3-4B for character generation and world narrative, so the model does more than answer questions: it becomes the engine driving the fantasy world's continuous evolution.
➤ Project link
🔗 https://huggingface.co/spaces/build-small-hackathon/aether-garden
🥉 Third prize | Tianwen reinterprets traditional Chinese fortune-telling with a 1B model
Tianwen is a self-reflection app that combines traditional Chinese fortune-telling with a small local model.
The system uses deterministic tools for calendar and birth-chart calculations, then has a LoRA fine-tuned MiniCPM5-1B generate gentler, more actionable interpretations.
The project also combines llama.cpp, local GGUF inference, safety filters, and rule-based fallbacks to explore how traditional cultural content can meet on-device AI.
➤ Project link
🔗 https://huggingface.co/spaces/build-small-hackathon/Tianwen
After 946 projects, the new exploration is just beginning
946 runnable applications, 263 MiniCPM projects, 74 teams that released custom models.
Behind these numbers is a concentrated exploration of where small models' capabilities end.
The point of small models is that they can be trained, controlled, and deployed for well-defined tasks, and ultimately find their way into someone's work, a real journey, or a fantasy world that keeps growing.
When models are small enough and tools open enough, the protagonists of AI innovation are no longer only labs with enormous compute.
They can also be every developer willing to start from a real problem and build the product with their own hands.
Build Small, Dream Big.
Congratulations to all the OpenBMB Awards winners, and thank you to every participant who built something with MiniCPM.
The next "small model, big build" is waiting for you to create it.