Tether & QVAC ran a hackathon in Lima with Peru's Ministry of Labor (@MTPE_Peru) and the Universidad de Lima (@udelima). The ministry asked students to build AI tools that help workers find a new job with the skills they already have. Over seven days, 78 students in 16 teams worked on the problem. The Minister of Labor, @JuanSheput, joined us for the award ceremony.

Paolo Ardoino (@paoloardoino) talked about QVAC at Plan ₿ Forum (@LuganoPlanB) last week. Watch him explain how QVAC is built on seven years of Tether technology, why intelligence sovereignty and privacy matter if AI ends up in the hands of a few companies, and why Tether is in a rare position to build QVAC. And remember: “Not your AI, not your intelligence.”

Why not just run everything in the cloud? "What if the connection drops and the robot cannot connect anymore to the data center? Will the robot stop?" @paoloardoino, CEO of Tether

Agent Emergence put QVAC inside an actual 2 GB camera, with detection on its NPU, VisionPsy-Nano describing the scene, and not a single frame leaving the device. They started from our open-source smart camera example and built on it to create this awesome real-life demo. Not only did they connect it to a real camera, they also optimised it by running detection on the neural processor and the analysis on the CPU. Brilliant.

The AI Yako hackathon kicks off this Saturday, 24 October, at Node NBO in Nairobi, and demo day is 31 October. Teams build AI that runs on their own device with QVAC, on three tracks: offline AI apps, moving a cloud product onto the device, and on-device agentic finance with WDK. Projects are judged with the internet switched off. Register here:

Should a few companies own all of AI? "Imagine how crazy it would be that two, three, four companies are going to handle all the intelligence for the entire humanity for the rest of the future." @paoloardoino, CEO of Tether

Why should AI be local first? "Humanity should invest in AI that is local first, that is owned by the people, that is available and inclusive to everyone." @paoloardoino, CEO of Tether

Half of Europe couldn't reach their bank app yesterday because one data center in London lost power. Now imagine the same outage taking down the AI your business depends on. Wild idea, but what if access to intelligence didn't need a data center?

Our PM can’t even take a flight without coming back with a QVAC prototype. This one turns any exam curriculum into practice tests, and it works without internet. Want to give it a try or build on it? It's accessible here:

QVAC SDK 0.22 is out, and it adds Laya, a model that makes decisions instead of writing text. We added: - Laya support: answers the questions you define, each with a probability - CUDA support: for LLM, embedding and VLA models. - MiniMax-H3 image to video: start a clip from your own picture - MOSS-SoundEffect v2: create sound effects on device - Model fit checks: broader model coverage, improved estimates for model combinations, and opt-in load refusal before model loading begins - Parallel tool calls: agents run several tools in one turn Learn more in the thread below. Complete release doc here: Star ⭐️ for updates.

We just released QVAC TranscribePsy, a speech-to-text model small enough to run inside a phone app. It covers 25 European languages on a laptop and 8 on a phone, and on an iPhone 17 Pro Max's Neural Engine the 8-language build transcribes a 10-minute recording in 1.65 seconds. Glasses, robots and phones are moving to voice, and each of them first has to write down what you said. Today that step mostly happens on a server. TranscribePsy does it on the phone or laptop, so dictation and meeting notes keep working with no signal, and a private conversation can be transcribed without the recording ever leaving the room. Run it on our qvac-fabric-speech.cpp's parakeet engine, which optimizes for efficient execution and delivers the highest performance. Get the models:

Our third llama.cpp PR in a week has been merged. This one, by @gianniCor, fixes llama.cpp's Vulkan backend, where choosing the next token could freeze an NVIDIA GPU. Thanks to @ggml_org for the review.

We optimized NVIDIA's Parakeet 0.6B models to transcribe twice as fast on the @Snapdragon 8 Elite, a flagship @Qualcomm phone chip. We moved the work from the chip's graphics processor to its AI processor, the Hexagon NPU. On a OnePlus 13s, Parakeet CTC 0.6B now transcribes a minute of audio in 970 ms, with the same text the GPU produced. Thanks to @NVIDIAAI for Parakeet and @ggml_org for the code we built on.