Maziyar Panahi

OpenMed

The open-source standard for clinical AI: 2,000+ Apache-2.0 models for de-identification, clinical NER, coding, and QA — built to run on-prem, in your VPC, on a Mac, or on an iPhone. 398M+ downloads on Hugging Face, 5,000+ GitHub stars, and the #1 most-referenced organization in Hugging Face's State of Open Source report (Spring 2026).

• openmed on GitHub — Python package, CLI, and MLX/Swift on-device engine; near-weekly releases, 50+ contributors
• OpenMed on Hugging Face — the model hub
• OpenMed NER — state-of-the-art on 12 public biomedical benchmarks
• OpenMed 2.0 — 56 languages, ~51 offline ID validators, and a k-anonymity release-risk workflow that refuses to auto-sign expert determinations

Next: OpenMed's own reasoning models — on-device, local-first — and multimodality across X-ray, pathology, and medical speech-to-text.

OpenMed Agent

The enterprise layer built on the open-source core — a terminal-native, inspectable clinical AI agent that is privacy-first, reviewer-gated, and running on iPhone. Early access is open at agent.openmed.life.

• OpenMed Agent — terminal-native clinical AI agent on the Hugging Face stack (early access)
• OpenMed Agent on iPhone — discharge-summary reader (on-device redaction of 13 identifiers + GPT-5.5 structured read)
• OpenMed Agent now runs on iPhone — “Patient Private Intelligence” (reads Apple Health, reasons with GPT-5.5, sends 0 identifiers)
• OpenMed Agent + Claude Opus 4.8 on contested aspirin guidance (declined to over-synthesize primary-prevention evidence)
• 16-step integrated workup demo (10,036 health records, GPT-5.5 via OpenAI Codex SDK)
• 14-step synthetic ED case (9/9 evidence domains, reviewer-gated artifacts)
• RADV audit walkthrough (Medicare Advantage chart review)

Welna

Know your numbers. Own your health story. Welna is my consumer iOS app, powered by OpenMed: it reads your Apple Health data, redacts identifiers on-device, and preps you for your next doctor's visit — your name, dates, and IDs never leave the phone.

• Welna: Understand Your Health — on the App Store

Posts

Real device, real input, visible result — demos over decks.

• 0% → 100% Sycophancy With One Vector: Replicating Anthropic's Emotion-Vectors Paper Across 7 Open-Weight Models
• Training mRNA Language Models Across 25 Species
• SynthVision: Building a 110K Synthetic Medical VQA Dataset with Cross-Model Validation
• OpenMed is the #1 most referenced organization powering open-source AI research on Hugging Face (State of Open Source report, Spring 2026)
• The ML Engineer's Guide to Protein AI
• From Golden Gate Bridge to Broken JSON: Why Anthropic's SAE Steering Fails for Structured Output
• Unlocking Healthcare AI: Releasing State-of-the-Art Medical Models for Free
• OpenMed: Six Months of Open-Source Medical AI and the Road Ahead

Research

• Arcee Trinity Large Technical Report
• OpenMed NER: Open-Source, Domain-Adapted State-of-the-Art Transformers
• INTELLECT-1 Technical Report
• Google Scholar

Models

600+ of these run natively on iPhone and Mac via MLX.

• Privacy Filter v2 (Nemotron, MLX 8-bit) — 755 tok/s streaming redaction on a Mac: 1,152 identifiers across 22 PII categories in a 13,000-token clinical file
• OpenMed model hub (healthcare-focused model ecosystem)
• 105 open-source PII detection models for Dutch, Hindi, and Telugu
• Personal Hugging Face model hub (general purpose, agentic, quantized model ecosystem)
• 20 open-source PII redaction models — Vietnamese, Bengali, Korean, Chinese (+ 20 MLX-native variants)
• 30 open-source PII models — Arabic, Japanese, Turkish (+ 30 MLX Swift packages)
• openmed v1.4.0 — Multilingual Privacy Filter family now first-class
• Multilingual Privacy Filter retrained on 1M+ samples across 16 languages
• First-ever fine-tune of OpenAI's Privacy Filter on NVIDIA's Nemotron-PII dataset
• MLX-native OpenAI Privacy Filter (24–33× faster) + GLiNER Clinical NER
• State-of-the-art medical NER on iPhone via Swift/MLX
• 35 open-source PII detection models for Portuguese
• OpenMed 1.0.0 — MLX backend + Swift package, models running natively on iPhone & Mac
• ~1 billion rows of psychiatric genetics data (PGC GWAS summary statistics) on Hugging Face

RL Envs

51 open medical RL environments on Prime Intellect — the same stack used to post-train Arcee's Trinity Mini for scientific tool-use.

About

I've spent 16 years in public academia: twelve at CNRS, where I'm a lifetime civil servant and run the AI & HPC infrastructure at ISC-PIF (140+ servers, 380B+ documents), after research-assistant and cloud-architect years at Multimedia University and the University of Malaya.

Before OpenMed, I spent seven years at John Snow Labs leading Spark NLP — 150M+ downloads and 100,000+ pretrained models and pipelines. Since 2024 I've also been in post-training at Arcee AI: GRPO, RL environments, and the INTELLECT-1 and Trinity technical reports.

The mission hasn't moved: build the default open engine for every hospital, pharma company, and public-health agency that needs a secure, sovereign medical-intelligence platform — and keep the models, the training, and their limits out in the open. The longer version is on LinkedIn.