What Does It Take to Put a 35B-Parameter Model on a Phone? Come See for Yourself
What does it feel like to fit a 35-billion-parameter multimodal MoE model onto a phone — one that can understand what the camera sees and answer questions about it?
At this year's GOTC, we'll bring live demos of our on-device multimodal models, a model experience app, and a fun interactive activity for GitHub developers. Come to the booth, try the models yourself, and talk with us about the real questions that come up from model deployment to application development.
Whether you're building on-device AI applications, exploring smart hardware, or simply want to learn about OpenBMB's open-source projects, you'll find something here worth seeing, trying, and talking about.
Booth Highlight #1: MiniCPM-V 4.7, live on an iPhone
When a multimodal model moves onto a phone, the world in front of the camera gains a new way to interact. At GOTC we'll demo MiniCPM-V 4.7 × iPhone Duo live. Come to the booth and watch how the model understands a scene and responds to questions — and feel what interaction looks like when multimodal capability meets a phone.
For developers working on on-device AI, this demo is also a concrete entry point for watching a model land in the real world: from capturing the frame to the model's response, is the interaction smooth? Can the model's performance meet the needs of the application? When device resources are limited, how should the model and the application share them? These questions often decide whether a model capability can become a product feature that users are willing to come back to.
On site, we'd also like to talk about the engineering behind the demo: how to choose a model for a given device, how to think about memory footprint and inference efficiency, and how to wire multimodal capability into a real application. If you're exploring AI for phones, IoT devices, or robots, bring your device constraints and your use case and let's talk. Starting from one live demo, we can discuss how your project might move forward.
Figure 1: MiniCPM-V 4.7 × iPhone Duo — a live multimodal demo
Booth Highlight #2: One app, every model capability
The most direct way to understand a model is to try it. With different kinds of tasks, hands-on experience also helps you judge which capabilities fit your application and which still need further validation. This time we're bringing an app that integrates all OpenBMB models, offering a single entry point to language, multimodal, speech, and document understanding capabilities — so you can experience the community's model results in one place. Start wherever you care most: see how the language model responds, feel the multimodal interaction, try the speech capabilities, or check how document understanding performs. You can also switch between directions and look for new possibilities for the application you're sketching out.
If you already have a clear need, connect what you experience with your own business scenario: does the model output meet expectations? Is the interaction pattern right for your target users? What adaptation and development work would integration require? And if you're just starting to explore AI applications, this single entry point can give you a more intuitive picture and point you toward directions worth pursuing.
We'd also love your feedback. Which capabilities make you want to dig deeper? Which experiences still have room to improve? What tools or support would you like the open-source community to add? Talk to us on site — let one trial become a more concrete development conversation.
Figure 2: Experience OpenBMB's model capabilities through a single app
Booth Highlight #3: What rank does your public GitHub footprint deserve?
What kind of report would you get if AI read the public open-source footprint you've left on GitHub? Come to the OpenBMB booth and try ghfind, an open-source scoring engine for GitHub developers. Enter your account, wait about 30 seconds to a minute, and you'll get a score for your public open-source footprint, a matching rank, and a playful AI write-up — plus a shareable card.
See how the report presents your public contributions, read what the AI has to say, and decide whether it got to the point. You can also try it with friends and share each other's cards, turning open-source experience into a topic of conversation. No GitHub account? You can still join in: pick a well-known developer's account and see how contributions you're familiar with get interpreted.
Behind the fun, ghfind also demonstrates an application pattern that combines rule-based computation with model understanding: scores are computed from public data and rules, while MiniCPM-V-4.7-35B-A3B understands the evidence of contribution and generates the report.
For developers, beyond collecting the card, there's more to discuss: how such applications organize evidence, generate reports, and make results easier to understand.
Bring your curiosity and take the test — see what rank your open-source footprint earns, and bring us your thoughts and suggestions about the report.
Figure 3: Test it, share it — let AI write a sharp take on your public open-source footprint
About OpenBMB
OpenBMB is an open-source community initiated by Tsinghua University's Natural Language Processing Lab (THUNLP) together with ModelBest. It is dedicated to bringing lightweight, efficient large models to on-device devices and smart hardware, making it easier for developers to turn model capabilities into real applications.
Centered on the MiniCPM family of on-device models, the community covers language, multimodality, document understanding, and speech, and also open-sources agent projects such as ChatDev, PilotDeck, and UltraRAG, along with tools for model training, compression, quantization, inference optimization, and data evaluation — building a full-stack open-source ecosystem that runs from model research to on-device deployment.
Today, OpenBMB models have been downloaded more than 70 million times, with over 170,000 GitHub stars and more than 22,000 Hugging Face likes. The MiniCPM series has repeatedly topped the GitHub and Hugging Face trending charts, drawing developers around the world to use, explore, and contribute.
"Bring large models to every home" is OpenBMB's vision — and the driving force behind our continued open-sourcing.
Come to the Booth and Talk About Your Use Case
At GOTC, we look forward to meeting developers working on on-device AI, multimodal models, document understanding, and agent applications — and to talking face to face with smart hardware vendors, open-source maintainers, and potential contributors.
How do you choose a model under tight memory limits? After quantization, how do you balance quality and speed? How do you bring multimodal capability onto existing devices? How should on-device and cloud workloads divide the work? How can open-source projects lower the barrier to use and contribution? Bring these questions to the OpenBMB booth.
Whether you're already deploying MiniCPM or just starting to explore on-device AI, come watch the demos, try the apps, and tell us about your devices, your projects, and your ideas.
See you at booth C21 at GOTC — let's turn model capabilities into more real-world applications.
⏰ November 14 (Saturday), 10:00–17:00
📍 Beijing · Tongming Lake Convention & Exhibition Center