Llama 3.2 Vision Instruct Medical Imaging Fine-Tune vs Perplexity Health

“The primitive here is a vision-language model with a domain-specific instruction fine-tune released as open weights — that's a real, nameable thing, and it matters. The DX bet is correct: drop the weights on Hugging Face under a research license so a team can pull them with one `transformers` call and run inference on-prem, which is exactly what hospital IT requires. The moment of truth is the first inference call with a DICOM-converted PNG — if the system prompt examples in the model card are solid, this survives the 10-minute test; if they're vague, researchers are on their own. My one gripe: the research license creates a hard fork from the permissive Llama community, so every downstream fine-tune has to re-negotiate terms, and that friction is a real DX tax.”

“Connecting 1.7M EHR providers via FHIR/API without building any hardware is exactly the right infrastructure play. If they open a developer API layer on top of this health data graph, every health app will want to plug in. The data moat here could be enormous.”

“Category is open-weight medical vision LLM; direct competitors are Google's Med-PaLM 2 and Microsoft's BiomedCLIP, both of which are closed or heavily gated — so Meta's move to open weights is genuinely differentiated, not just marketing. The scenario where this breaks is any real clinical deployment: the research license explicitly forbids diagnostic use, so the addressable user is a researcher with GPU access, not a radiologist. What kills this in 12 months is not a competitor but regulatory clarity — if the FDA signals that research-licensed models can't touch real patient workflows even in research contexts, the use case shrinks to benchmarking papers. What would have to be true for me to be wrong: the research community uses this to produce fine-tunes that actually hit FDA breakthrough device designation, which is plausible but not a given.”

“Perplexity has had data sourcing controversy before. Trusting them with your EHR and biometric data is a much higher-stakes bet than trusting them with web search. One breach, one data-sharing revelation, and the regulatory blowback would be severe — HIPAA exposure is no joke.”

“The thesis here is falsifiable: within three years, medical AI will be dominated by institution-hosted open-weight models rather than API-dependent closed ones, because HIPAA and international data-residency rules make cloud inference a liability, not a feature. The dependency that has to hold is that GPU costs continue falling fast enough that a mid-sized hospital system can afford to run a 90B-parameter model on-prem — that trend line is real and on-time. The second-order effect nobody is talking about: this shifts the center of gravity in medical AI from a handful of well-funded startups with proprietary model access to radiology departments and academic medical centers with compute budgets, which democratizes the research surface but also fragments quality benchmarks. The future state where this is infrastructure is a world where every major health system has a model registry the way they have a formulary — and this release accelerates that norm.”

“Longitudinal personal health AI is the killer app that makes everyone a power user of their own data. When you can ask 'why was my HRV tanking in February?' and get a real answer, health AI stops being aspirational and starts being essential. Perplexity just claimed the territory.”

“The buyer here is a clinical researcher or academic institution, which means the check comes from a grant budget or a research IT line — small, slow, and heavily committee-gated. Meta isn't building a business with this release; they're publishing a research artifact, so the 'pricing is free' observation misses the point — the real question is what Meta captures, and the answer is talent signaling and ecosystem influence, not revenue. The moat for anyone trying to commercialize on top of this is essentially nonexistent: the weights are public, the fine-tune recipe will be replicated, and the research license strips out the highest-value commercial use cases. If I were a founder building on this, I'd need a very specific workflow integration — structured report templating, PACS system connectors, audit logging — to create switching costs, because the model itself is not the business.”

“Generating pre-appointment summaries I can actually share with my doctor is the kind of practical health feature I've wanted for years. The UX of 'ask a question, get your data back in plain language' is dramatically better than digging through the Health app graphs.”