This private repository contains an OpenMed MLX-LM conversion of kogai/laneformer-2b-it. It is packaged for local Apple Silicon text generation through OpenMed's Python MLX interface and mlx-lm.

At a Glance

Field	Value
Source model	`kogai/laneformer-2b-it`
MLX repo	`OpenMed/laneformer-2b-it-q4-mlx`
Task	Text generation
Runtime	Python `openmed[mlx]` / `mlx-lm`
Quantization	4-bit affine, group size 64
Parameters	2.32B
Source revision	`b4f40adc413c2c5268ab89cf666ade37148d8d4b`
License	Custom upstream license, see source license link

OpenMed MLX Status

Python MLX: supported through openmed.generate_text(...) and openmed.mlx.OpenMedMLXLanguageModel.
Swift MLX: not supported for this causal language model artifact. Swift OpenMedKit MLX currently targets OpenMed token-classification artifacts.
Privacy posture: this artifact is intended for local inference. Do not send protected health information to hosted demos or external services.
Safety posture: OpenMed does not treat this model as a medical device and does not auto-trigger clinical decisions.

Use This MLX Snapshot

bash

hf download OpenMed/laneformer-2b-it-q4-mlx \
  --local-dir ./laneformer-2b-it-q4-mlx

Python Quick Start

pip install "openmed[mlx]"

python

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 when you want this preconverted MLX artifact explicitly. OpenMed also accepts kogai/laneformer-2b-it and laneformer-2b-it as compatibility aliases that resolve to this private OpenMed artifact.

Use This Preconverted MLX Repo Directly

python

from openmed.mlx import OpenMedMLXLanguageModel

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

You can also load this directory directly with mlx_lm.load(...).

Artifact Notes

Format: MLX-LM model directory.
Weights: model.safetensors.
Custom model implementation: laneformer.py, referenced by config.json through model_file.
Tokenizer assets: tokenizer.json, tokenizer_config.json, special_tokens_map.json, and chat_template.jinja.
Quantization metadata is stored in config.json as 4-bit affine with group size 64.

CPU vs MLX Smoke Test

The private export verification used a 13-token prompt on Apple Silicon:

Runtime	Mean prefill	Tokens/sec
PyTorch CPU	0.3112 s	41.77
MLX q4	0.1201 s	108.21

Measured speedup: 2.59x for prefill on the smoke prompt. The CPU and MLX top-5 next-token sets overlapped on 4 of 5 token ids.

OpenMed/laneformer-2b-it-q4-mlx

Laneformer 2B Instruct q4 MLX for OpenMed