Using AI to help kids with basic education is a very worthy goal!

And here I thought that first batch of Hiraia.org phones was a lot to handle. This next batch of donated devices is so big that I'm gonna have to build an Android device provisioning pipeline to set them up automatically ... otherwise I'll be sitting here tapping screens all week. These are enough phones for my next 2 pilot school cohorts, the first of which I'm meeting up with this week! Big thanks to @gabusch and @YGGAlerts for making this happen so quickly ❤️🇵🇭 @qvac @tether

1) Hiraia isn't a flashy, exciting AI application (you'll never see any of its actual beneficiaries talking about it online) but I'd argue that it's more necessary now than ever, so these little bits of coverage means a ton ❤️ 2) That AI Summit sounds like a blast! Hopefully I'll have more flexibility to attend next year and share what I'm doing with Hiraia in person 🌞 #AISB26 @qvac @tether

Images from this morning's Hiraia launch at Calapacuan Elementary School in Zambales, Philippines ❤️🇵🇭 Possibly the first decentralized AI education app deployed with an academic institution on this side of the world, and it's built with the @qvac SDK and supported by a grant from @tether. Now the real work begins! Download the student and teacher app combo at Hiraia.org and pls follow me if you wanna check out our progress over this first semester. @paoloardoino @Yummyfunduk @shawnyao @gladstein

Really maximizing my hotel room the night before my first Hiraia pilot launch! @qvac @tether

QVAC SDK 0.20 is live. 🚀 We added support for: - ABot-World: generates a walkable world from a single photo - MiniMax-H3: create video with synchronised soundtrack - Video models can now run on GPUs with less VRAM - TranslatePsy-AfriSLM: translate between English & 19 African languages on-device - ACE-Step can now generate entire songs from a single sentence & more new features We also merged the latest changes in llama.cpp into QVAC Fabric, to ensure continuous compatibility. Learn more in the thread below. Complete release doc here: github.com/tetherto/qvac/rel…

A whole lotta phones to install Hiraia on!

We solved financial inclusion with USD₮. Now we are tackling intelligence inclusion with QVAC. 🤖 Join me for a Reddit AMA on October 7th at 3pm UTC to talk about on-device AI runtimes, privacy, local inference, and the future of decentralized tech.

Congratulations to Luis and to everyone at Calapacuan Elementary. Local AI is not only about privacy. It has the potential to improve lives and to reach people who would otherwise have no access to artificial intelligence at all. Luis is doing exactly that with Hiraia: a tutor that lives on the device and works where the connection does not.

Cross-posting this from Calapacuan Elementary Principal Marla Magat's FB page. Super close to our first pilot roll-out now! Check us out at hiraia.org 🇵🇭🇵🇭🇵🇭 @qvac Calapacuan Elementary School Makes History as First Pilot Site in the Philippines for Groundbreaking Offline AI Science App, HIRAIA. It is a major milestone for Philippine public education and digital equity, Calapacuan Elementary School has been chosen as the very first pilot testing site in the country for HIRAIA, an innovative, offline-first artificial intelligence science learning platform. Developed by prominent Filipino tech pioneer Luis Buenaventura, HIRAIA addresses one of the country’s most persistent educational hurdles: the digital connectivity divide. Backed by a ₱1-million research grant from the @Tether Foundation, the open-source initiative brings state-of-the-art AI tutoring directly onto entry-level mobile devices. Because the intelligence engine, science curriculum facts, and illustrations live entirely on the device, students can learn, explore, and review lessons without needing an active internet connection or mobile data subscription. Through this landmark deployment, Calapacuan Elementary School will receive a dedicated fleet of mobile learning gadgets pre-installed with the HIRAIA app. Aligned with the Department of Education’s MATATAG curriculum, HIRAIA supports interactive science concepts, dynamic flashcards, and quizzes in both English and Tagalog. The platform empowers elementary learners to master science at their own pace—inside the classroom or at home—while completely protecting pupil privacy by operating without ads, tracking, or account sign-ups. “The future of learning should not stop where the internet ends,” the school administration shared. “Calapacuan Elementary School is immensely proud to lead the nation as the pioneer testbed for HIRAIA. This partnership shows what happens when visionary Filipino tech leadership collaborates directly with grassroots public schools to put world-class learning tools in the hands of our children.”As the inaugural cohort of young scientists at Calapacuan Elementary School begins exploring HIRAIA, the data and classroom feedback gathered will play a vital role in fine-tuning the platform before its wider rollout across other grade levels and other schools nationwide.

🚨 A World Bank research showed that the average 15-year-old Filipino student possesses the math and science capabilities of a 10-year-old in Singapore, which means local public school children are up to five full school years behind their regional peers. For Luis Buenaventura (@helloluis), this is "heartbreaking," so he decided to build an AI tutor that works directly on entry-level mobile hardware: Hiraia, an open-source AI science tutor that runs completely offline. 👇 👉 Full Details: bitpinas.com/go/luis-hiraia

The first three Hiraia test devices are working pretty well! Can't wait for the big shipment next week ❤️🇵🇭

The most obvious use is as an intelligent customer-support triage but there would be interesting applications elsewhere. Theoretically this is like enjoying output of an ML or RNN pipeline without all the preprocessing.

Today you need less than $500 of hardware to locally run a model that outperforms models that were the best in the world 6 months ago Welcome to the future

Hey whenever you all are ready to start training sovereign LLMs, just lemme know

Visited Calapacuan Elementary School in Zambales this week and got to speak with the principal and the division superintendent. Crazy numbers: this public school has over 1,500 students and their ANNUAL working budget is just $40,000. The plan we agreed on is to do a limited pilot run of my AI science app #Hiraia with a single class of 35 students over the next 2 months to test its effectiveness as a grade school tutor. ❤️🇵🇭 Super excited to get this going so quickly and grateful for the support from @qvac and @tether @paoloardoino

To borrow from the great Phil Zimmerman: If open models are outlawed, only outlaws will have open models

How to tell that I'm having a particularly deep vibe coding session: I didn't notice that I'd been here for over an hour.

Hiraia Development Notes: I'm entering into a tricky quality-assurance phase in preparation for my pilot version 0.5. ICYMI what I currently have is an on-device AI tutor for Filipino public school kids, carrying 50,000+ grade school science facts and about 33,000 quiz cards. Since I am maintaining English, Tagalog, and Cebuano versions of each, that's over 150k facts and about 100k quiz cards. The fact bank is generally ok, but the quiz cards need to be audited for quality and accuracy. I've noticed some awkwardness in the Tagalog wording, and I don't think it's acceptable to launch this as is. The biggest challenge is that none of the open-weights models (DeepSeek, Kimi, GLM, Qwen, MiniMax) are strong enough to recognize good Tagalog and Bisaya translations consistently. I did several comparative tests last night, and their average error rate was somewhere between 25% to 50%, meaning they make mistakes on about 25% of the cards that they audit and rewrite. (In case you're wondering, Qwen 3.8 Max 0902 was the lowest error rate, and MiniMax M3 was basically just guessing the whole time. DeepSeek 4.1 and GLM 5.3 were very similar, and Kimi K3 was right in the middle.) Annoyingly, the mistakes were never the same 25%, so if you ran the same batch again, you would find errors in different cards. In other words, I couldn't trust the consistency of their output. So instead I'm forced to audit+rewrite using closed models, which as you can imagine is super expensive. My cost estimate for a GPT Sol-powered run was in the $2,000 range ASSUMING it got everything right in a single pass. I didn't even bother estimating with Fable. For obvious reasons this is entirely out of my budget range -- I could buy 20+ mobile phones for our pilot classroom with that much money. The one shining light here is Google Gemini, which has shown competitive translation capabilities (I assume 20 years of Google Translate has helped them a bit here) and is on some kind of (temporary?) sales promotion this month for Gemini 3.8 Flash. Google has a batching service called Gemini Flex that allows you to submit 5000 requests at a time and gives you a 50% discount. The tradeoff is that they will process your request lazily over a 24-hour period, so you don't know exactly when you'll get it back. This is perfectly fine with me given how much I could save with them: my estimate is that the total spend will be at less than $400 by tomorrow afternoon. I have over 60,000 quiz cards to audit, rewrite, audit, rewrite, over and over until I've gotten the error ratio to single digit percentages. Once we're done with that, I'll do a final APK rebuild and launch version 0.5! You can follow my progress at Hiraia.org 🇵🇭🇵🇭🇵🇭 @qvac @tether