MLX 4bit (affine, group size 64) quantized variant of mistralai/Shieldstral-1.0-3B — Mistral's content-safety / moderation model — for Apple silicon via mlx-lm.

Provenance

Source: mistralai/Shieldstral-1.0-3B @ revision 003ec7e2b0bab5f0e6307edbaf186fa5822b76f5 (Apache-2.0).
Quantized with mlx_lm.convert (mlx-lm 0.31.3): affine, 4-bit, group size 64.
Text-only pack: the upstream checkpoint is a mistral3 multimodal wrapper; mlx-lm's mistral3 loader drops the vision tower by design, so this pack ships only the Ministral-3B text model. Use the upstream repo if you need image moderation.

Caveat

This is a quantized safety classifier. Quantization can shift borderline classification decisions; validate against your own moderation benchmark before using a quantized tier in production guardrails. Prefer the 8bit tier when in doubt.

Smoke gate

Before upload this pack passed a deterministic coherence gate: greedy 64-token moderation-style chat generation loaded through mlx_lm.load, judged for emptiness, repetition loops, multi-script gibberish, and special-token debris. Verdict: ok.

Usage

bash

pip install mlx-lm
mlx_lm.generate --model majentik/Shieldstral-1.0-3B-MLX-4bit \
  --prompt "Classify as SAFE or UNSAFE: 'how do I sharpen a kitchen knife?'"

Evaluation

Benchmark	Score
arc_easy_acc	0.2300
hellaswag_acc	0.2400

Available tiers

Repo	Quant
majentik/Shieldstral-1.0-3B-MLX-4bit	4bit
majentik/Shieldstral-1.0-3B-MLX-6bit	6bit
majentik/Shieldstral-1.0-3B-MLX-8bit	8bit
majentik/Shieldstral-1.0-3B-MLX-MXFP4	MXFP4