Breaking

OpenMed's Local PII Redactor v2 Streams 755 Tokens per Second on a Mac

July 5, 2026 at 00:45 EDT

Healthcare
Open Source
Foundation Models

OpenMed's privacy filter v2 — a Nemotron fine-tune quantized to MLX 8-bit — was shown processing a 13,000-token clinical document entirely on a Mac, detecting and redacting all 1,152 identifiers across 22 PII categories. It ran at 755 tokens per second, with no data leaving the machine.

Open-Source · On-Device PHI Redaction

OpenMed redacts 1,152 identifiers in a clinical doc — 100% on your Mac

A privacy-filter model, fine-tuned on NVIDIA's Nemotron-PII and 8-bit quantized for Apple's MLX, streams a 13k-token medical record at 755 tokens/sec while stripping PHI — with no data ever leaving the device.

755

tokens/sec on a 13k-token document

1,152

identifiers detected across 22 categories

55+

supported entity types

100%

local — no cloud API, no data leaves device

Detection coverage grew nearly 7× — v1 → v2

Each block = a stacked share of covered categories. v2 covers the 18 HIPAA Safe Harbor identifiers and more.

v1 categories

55+

v2 categories

Names · dates of birth · SSN · credit cards · IBAN · MRN · blood type · IP addresses · API keys

8-bit quantization: ~1.7× faster, 47% smaller

Model size

2.6 GB

BF16

1.4 GB

8-bit (MLX)

1.7× faster than BF16

100% argmax agreement in 10-sample check

Where it shines

Local medical-record processing, no cloud
OCR + redaction in an iOS scanning app
FHIR output pipelines
Climbed 6 places over base model on the PII masking leaderboard

Still to verify

Non-Apple hardware falls back to slower PyTorch build
Large batch processing needs testing
Non-English multilingual use unproven
Minor quantization differences possible

Accuracy of the Nemotron fine-tune

0.9533

Macro B-F1

0.9910

Token accuracy

46 / 55

labels at F1 ≥ 0.90

Apache-2.0

fully open source

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.