MLX Backend (Apple Silicon)¶

OpenMed v1.5.5 expands native Apple Silicon acceleration via Apple MLX, including preconverted Arabic, Japanese, and Turkish PII token-classification artifacts.

That MLX story now has two surfaces:

• Python MLX through openmed[mlx] on Apple Silicon Macs
• Swift MLX through OpenMedKit on Apple Silicon macOS and real iPhone/iPad hardware

Installation¶

# From the repository root
pip install -e ".[mlx]"

This installs mlx, mlx-lm, huggingface-hub, transformers, tokenizers, safetensors, and Pillow.

Quick Start¶

from openmed import analyze_text
from openmed.core.config import OpenMedConfig

# MLX is auto-detected on Apple Silicon — no config needed
result = analyze_text(
    "Patient John Doe, DOB 1990-05-15, SSN 123-45-6789",
    model_name="pii_detection",
)
print(result.entities)

Privacy Filter decoding and compilation¶

The Python MLX Privacy Filter keeps eager execution as the default. Kernel compilation is an explicit performance experiment because fused kernels can produce unit-in-the-last-place floating-point differences:

from openmed.mlx.inference import PrivacyFilterMLXPipeline

pipeline = PrivacyFilterMLXPipeline(
    "/path/to/openai-privacy-filter-mlx",
    compile_forward=True,
)

Alternatively, set OPENMED_MLX_COMPILE=1. The values 1, true, yes, and on are accepted case-insensitively. An explicit compile_forward=True or compile_forward=False always overrides the environment variable. Leaving the argument unset and the variable absent selects eager mode.

BIOES transition tables retain only the most recent bias configuration, so varying biases cannot grow an unbounded matrix cache. The NumPy decoder and its pure-Python fallback reject NaN and positive-infinite emissions, state scores, or biases; negative infinity remains the supported representation for an impossible emission. Confidence reconstruction is rounded to binary32, matching the previous MLX probability tensor. Labels, text, and offsets must match exactly; allow an absolute score tolerance of 1e-6 across MLX devices or compiled/eager kernels.

Reference latency measurement¶

The following steady-state measurement was collected on August 23, 2026, on an Apple M2 Max MacBook Pro with 96 GB RAM and macOS 26.5.2. Both revisions used Python 3.11.10, MLX 0.32.1, NumPy 1.26.4, tiktoken 0.14.0, eager mode, and the same locally cached OpenMed/privacy-filter-mlx artifact at revision 833fa7ea3fd36148900deea2d55bdadd4b90efa9. Model loading was excluded. Each cell is the median of 12 synchronized pipeline calls after three warmups over synthetic text encoded to the exact sequence length.

“Before” is upstream revision 55232ecf; “Optimized” is the repaired #2946 working tree based on 61e318b2. To reproduce the method, load one pipeline per revision, generate text by decoding the first N tokens of a repeated synthetic sentence and assert that re-encoding returns exactly N tokens, run each batch and sequence-length pair three times, then time 12 further calls with time.perf_counter(). A synthetic person/email/phone fixture produced identical entity JSON and confidence values before and after the optimization.

Python MLX-LM Quick Start¶

OpenMed also exposes MLX-LM causal language models through the same openmed[mlx] extra. The first supported model is the private OpenMed/laneformer-2b-it-q4-mlx conversion of kogai/laneformer-2b-it.

from openmed import generate_text

response = generate_text(
    messages=[
        {
            "role": "user",
            "content": "Explain why local clinical language models matter.",
        }
    ],
    model_name="OpenMed/laneformer-2b-it-q4-mlx",
    max_tokens=128,
)
print(response)

Use OpenMed/laneformer-2b-it-q4-mlx to request the preconverted OpenMed MLX artifact explicitly. The resolver also accepts these aliases:

• kogai/laneformer-2b-it
• laneformer-2b-it
• a local directory containing the converted MLX-LM artifact

For explicit reuse across several prompts, keep the model loaded:

from openmed.mlx import OpenMedMLXLanguageModel

runner = OpenMedMLXLanguageModel("OpenMed/laneformer-2b-it-q4-mlx")
print(runner.generate("Define delayed tensor parallelism.", max_tokens=128))

Python Vision-Language Quick Start¶

OpenMed includes a native Cohere Compass runtime for the North Micro Vision MLX family. It consumes data-only model artifacts directly:

from openmed.mlx import OpenMedMLXVisionLanguageModel

model = OpenMedMLXVisionLanguageModel(
    "OpenMed/North-Micro-Vision-Instruct-6bit-mlx"
)

text = model.generate(
    "Explain why local processing can improve clinical-document privacy.",
    max_tokens=96,
)
document = model.generate(
    "List the visible medication and dose.",
    image="synthetic-clinical-note.png",
    max_tokens=128,
)

The first load downloads the selected artifact from Hugging Face. Pass a local model directory instead for an air-gapped deployment. Images and prompts are processed locally; OpenMed does not add telemetry or a cloud inference fallback. Treat generated text as sensitive and validate it against the source image before consequential use.

The compatible repositories use one naming family:

• OpenMed/North-Micro-Vision-Instruct-4bit-mlx
• OpenMed/North-Micro-Vision-Instruct-5bit-mlx
• OpenMed/North-Micro-Vision-Instruct-6bit-mlx
• OpenMed/North-Micro-Vision-Instruct-8bit-mlx
• OpenMed/North-Micro-Vision-Instruct-bf16-mlx

Maple Preview: structured clinical tasks¶

OpenMed resolves maple, maple-preview, and the original deepgrove/maple-preview id to the pinned deepgrove/maple-preview-2bit-mlx artifact. The wrapper supports validated PII spans and redaction, clinical entities, directed relations, and note-grounded answers:

from openmed import MapleClinicalAssistant, MapleTask

assistant = MapleClinicalAssistant("maple")
note = "Synthetic patient Alex Doe takes metformin 500 mg daily."

masked = assistant.complete_task(MapleTask.PII, note)
relations = assistant.complete_task(MapleTask.RELATIONS, note)
answer = assistant.complete_task(
    MapleTask.REASONING,
    masked.redacted_text or "",
    question="Which medication and dose are documented?",
)

The model is loaded lazily. Custom executable model code is pinned to an immutable Hub revision, and the task parser rejects out-of-bounds spans, unknown relation endpoints, extra schema keys, and malformed JSON. See Maple Preview on device for native iOS, Android, browser, and mixed-bit export workflows.

Paged KV Cache for Long Notes¶

Long clinical-note prompts can opt into OpenMed's paged KV-cache planning for MLX-LM generation. The plan uses a fixed page pool, chunked prompt prefill, and a sliding in-memory window when a prompt exceeds the configured budget:

from openmed.mlx import OpenMedMLXLanguageModel, PagedKVCacheConfig

runner = OpenMedMLXLanguageModel("OpenMed/laneformer-2b-it-q4-mlx")
cache = PagedKVCacheConfig(
    memory_budget_bytes=512 * 1024 * 1024,
    page_size_tokens=128,
    chunk_size_tokens=512,
    # Set this from the loaded model's KV footprint when known.
    bytes_per_token=65_536,
)

response = runner.generate(
    long_note_prompt,
    max_tokens=256,
    paged_kv_cache=cache,
)
print(runner.last_paged_kv_cache_plan.to_dict())

The exact dense-cache context supported by a budget is:

floor(memory_budget_bytes / (page_size_tokens * bytes_per_token)) * page_size_tokens

For example, with 128-token pages:

Prompts at or below that exact context keep byte-identical generation inputs to the dense-cache path while using chunked prefill. Longer prompts degrade gracefully by bounding resident KV pages to the configured window and recording the older tokens that require recompute/eviction accounting; tokens inside the resident window remain exact.

To force a specific backend:

config = OpenMedConfig(backend="mlx")   # Force MLX
config = OpenMedConfig(backend="hf")    # Force HuggingFace/PyTorch
config = OpenMedConfig(backend=None)    # Auto-detect (default)

How It Works¶

1. Auto-detection: On Apple Silicon Macs with mlx installed, OpenMed automatically selects the Python MLX backend.
2. Artifact packaging: Supported conversions now produce a self-contained MLX artifact with:
3. openmed-mlx.json
4. config.json
5. id2label.json
6. tokenizer assets
7. weights.safetensors by default
8. weights.npz as a fallback when needed
9. Shared contract: That same MLX artifact shape is now the contract for both Python MLX and Swift MLX.
10. Identical output shape: MLX produces the same entity format as the HuggingFace backend, so downstream entity merging and PII handling stay consistent.

The public runtime focuses on automatic preparation at first use. OpenMed's broader cross-architecture conversion work is still being generalized privately across the full model collection.

Quantized export certification is documented in MLX Quantized Export Certification, including INT4 recall-delta reports and the certified manifest field.

Architecture Coverage¶

As of May 4, 2026, the current public MLX path covers these families:

• bert
• distilbert
• roberta
• xlm-roberta
• electra
• deberta-v2 / DeBERTa-v3-backed experimental GLiNER-family artifacts
• openai-privacy-filter
• privacy-filter-nemotron / privacy-filter-multilingual artifacts through the OpenAI Privacy Filter runtime
• cohere-compass for native-resolution text-and-image generation through OpenMedMLXVisionLanguageModel

Python MLX and Swift MLX now share the same artifact contract for OpenMed PII, Privacy Filter, OpenAI Nemotron Privacy Filter, OpenMed Multilingual Privacy Filter, and experimental GLiNER-family tasks. The Arabic/Japanese/Turkish PII rollout adds 28 supported -mlx repos now; unsupported ModernBERT, Qwen3, and Longformer PII checkpoints remain deferred until those architectures land in the converter.

MLX-LM causal text generation is a separate Python-only artifact contract. It uses MLX-LM files such as model.safetensors, tokenizer assets, config.json, and custom model_file implementations when needed. Laneformer support is available through OpenMed/laneformer-2b-it-q4-mlx. Maple Preview support uses the pinned deepgrove/maple-preview-2bit-mlx custom MLX-LM artifact and a native OpenMedKit implementation for Swift. Cohere Compass vision-language generation instead uses the shared OpenMed Python/OpenMedKit contract described above.

Architectures still in active rollout:

• modernbert
• longformer
• eurobert
• qwen3

That rollout is about making the converter universal and repeatable across the whole OpenMed collection, not just a single pilot checkpoint.

Fallback Behavior¶

If MLX is not available (non-Apple hardware, or mlx not installed), OpenMed automatically falls back to the HuggingFace/PyTorch backend. No code changes required.

That automatic fallback applies to the token-classification backend. MLX-LM text generation and the Compass vision-language runtime require Apple Silicon, the openmed[mlx] dependencies, and a supported OpenMed MLX artifact.

MLX and Swift Apps¶

OpenMedKit can now load supported OpenMed MLX artifacts directly in Swift.

• Use Python MLX when you are running OpenMed from Python on Apple Silicon.
• Use Swift MLX when you want the same supported MLX artifact to run in an Apple app on:
• Apple Silicon macOS
• a real iPhone/iPad device
• Use CoreML when you already have a bundled Apple model package or need a fallback path outside Swift MLX.

Swift MLX does not target iOS Simulator.

Swift MLX Quick Start¶

import OpenMedKit

let modelDirectory = try await OpenMedModelStore.downloadMLXModel(
    repoID: "OpenMed/OpenMed-PII-ClinicalE5-Small-33M-v1-mlx"
)

let openmed = try OpenMed(
    backend: .mlx(modelDirectoryURL: modelDirectory)
)

let entities = try openmed.analyzeText(
    "Patient John Doe, DOB 1990-05-15, SSN 123-45-6789"
)

See OpenMedKit (Swift) for the full Swift setup flow.