OpenMed

Just got GLM-5.2 to run an entire clinical case by itself, agentically 🔥 it plans, calls tools, and writes the disposition. Gemma 4 pulls the structure, OpenMed does the de-id, all open models on Hugging Face And I won't name names, but a healthcare AI vendor just leaked 1.4M patients this month. Names, SSNs, diagnoses. Because the data sat in someone's cloud. Here OpenMed strips every identifier on-device first, so GLM-5.2 only ever sees redacted text and never learns a single name. The raw note never leaves the machine. Why does yours still see it?

I finally added the thing everyone kept asking for: OpenMed models now run in your browser and on Android, via ONNX. 🔥 Watch it de-identify a clinical note fully on-device: 19 identifiers gone, 0 bytes leaving the machine. No server, no API. What should run on-device next?

I believe the releases of GPT-5.6 Sol and Fable 5 mark another major milestone, perhaps even a true turning point, in the age of AI. It feels similar to the shift we experienced six or seven months ago, when agents began to rise and AI started taking over a large share of real-world coding. But this moment feels even bigger. These models no longer feel like impressive experiments or tools with obvious gaps. They feel complete, deeply capable, reliable, and astonishingly good. The quality, reasoning, coding ability, autonomy, and overall coherence have reached a level that makes the one feel that age of AI have now grown up to maturity. Of course there is much much more room to grow in intelligence and capabilities. But this is one of those moments when progress does not feel incremental. It feels like a new era has begun. AI models are just insanely good now, and the possibilities feel almost limitless!

I finally got GPT-5.6 Sol running inside OpenMed Agent on max, on a 186-step clinical marathon. Such a MASSIVE prompt. 🔥 132 steps in order, 15 workflows, and it found +0.475 RAF in upgrades then refused to bill them without a human. Long-horizon clinical AI is getting real!

I redacted a patient chart. Then I copied the name right back out from under the black box. 🔥 Most "redacted" PDFs never scrub the text layer. So openmed 1.8 ships verify-pdf: it fails your redaction unless the text is actually gone. It's out! What should it catch next?