On-device AI and DevRel @ swmansion | React Native ExecuTorch | Private Mind | Velonos
Kraków
Joined February 2025
- Tweets211
- Following145
- Followers263
- Likes500
Pinned Tweet
I built a 3D scanner in React Native that runs fully on-device :D
You take 8 photos of an object or scene, Depth Anything 3 turns them into 1.2 million points you can spin with your finger, ~3s on an iPhone 16 Pro. The model even figures out where the camera was standing for each photo.
Watching @theworldlabs drop their Atlas world model two days ago sent me down this path. That's spatial intelligence at cloud scale and I wanted to see how much of it already fits on your phone.
stack: @swmansion's react-native-executorch + typegpu
As always @liquidai doesn't disappoint.
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
I’ve built an app that runs Meta’s SAM 3 (848M params) fully on my iPhone :D
Had a few free hours while my agent was working, so I decided to export a big model and make it run on the phone as efficiently as possible. I love computer vision, so it had to be SAM 3.
Type a word and it segments every instance of it: 16 boats, 12 houses. The photo is encoded once (~2s), then each new word takes ~0.25s, so the masks update as you type. All of it works in airplane mode.
Expo SDK 58 users, this one's for you. If your Android build broke after upgrading, bump to 0.10.6 :D
react-native-executorch 0.10.6 released!
React Native 0.87/0.88 and Expo SDK 58 support, correctly normalized CLIP image embeddings, and Vulkan fixes.
swmansion.com/changelog/reac…
The new Claude Code desktop app is so much better. Running multiple sessions on different features side by side finally feels natural.
Also the animation on an empty session is weirdly satisfying.
Great job @ClaudeDevs 👏
We also have a brand new landing page so check it out:
privatemind.swmansion.com/
We shipped web search in Private Mind. A hidden WebView browses for you, and on-device models find the useful parts of each page. No search API, no server in between.
Links in the thread 👇
Btw huge kudos to @Krzysztof2236 for his work on this feature.
We shipped web search in Private Mind. A hidden WebView browses for you, and on-device models find the useful parts of each page. No search API, no server in between.
Links in the thread 👇
Btw huge kudos to @Krzysztof2236 for his work on this feature.
I’ve built a Funko Pop scanner for my reactnativeCon presentation :D
It takes only 44ms from the scan to a result, fully on-device. Thanks to the improvements in the latest release of react-native-executorch I was able to implement it only by writing TypeScript code.
Blog about using agents for the ML part coming soon.
Code and models: github.com/NorbertKlockiewic…
Huge, huge, huge!
Software Mansion is now a Select partner in the Claude Partner Network 🎉
If you're looking for an AI-first team that builds with the latest tools, let's talk.
More in our blog post: swmansion.com/blog/software-…
@AnthropicAI @claudeai
No better place to polish a talk than a train to the conference. See you all at @reactnativecon
I can't believe how good working with Opus 5.5 feels. This model got rid of all the problems from previous version. I finally understand what it says and it's very precise when doing it's task without bloating the codebase. And I don't use 1/4th of my weekly limit in 3 hours.
Anthropic really cooked with this one, excited for Fable 5.5.
I'll be in Berlin on 7 and 8 October for @reactnativecon, where I'm giving a talk on training and deploying custom ML models in React Native, and skipping the hard parts by handing them to agents.
It's going to be 🔥
Norbert Klockiewicz retweeted
Introducing ruNNtime v0.1.0 🧠
We built a fast way to run neural networks straight in your browser, powered by our very own TypeGPU.
Guys did incredible job with this release, whole library was rewritten to make creating custom pipelines easier. Besides a huge performance upgrades and my favourite part: running models on GPU on android with vulkan.
Look at the sticker on his belly 😍
We asked Reachy Mini by @huggingface & @pollenrobotics what it thinks of on-device AI. 🤖
Reachy hears with Whisper, speaks with Kokoro (both running locally), while Claude does the reasoning. ⬇️