Ornstein-Hermes-3.6-27B SABER MLX 8-bit

This is an MLX 8-bit quantized conversion of GestaltLabs/Ornstein-Hermes-3.6-27b-SABER.

The source model is a SABER-edited release candidate of GestaltLabs/Ornstein-Hermes-3.6-27b, selected for its observed refusal/KLD tradeoff on the source evaluation set. See the source model card for the full methodology, evaluation details, attribution, and limitations.

Quantization

field	value
Format	MLX safetensors
Quantization	affine
Bits	8
Group size	64
Reported bits per weight	8.501
Uploaded storage	~28.6 GB

Use With MLX

Install mlx-lm, then load this repo directly:

pip install -U mlx-lm

mlx_lm.generate \
  --model GestaltLabs/Ornstein-Hermes-3.6-27b-SABER-MLX-8bit \
  --prompt "Explain quantum computing in simple terms."

Python:

from mlx_lm import load, generate

model, tokenizer = load("GestaltLabs/Ornstein-Hermes-3.6-27b-SABER-MLX-8bit")
response = generate(
    model,
    tokenizer,
    prompt="Explain quantum computing in simple terms.",
    max_tokens=256,
)
print(response)

For chat-style usage, apply the tokenizer chat template used by the source model before generation.

Notes

• This repo contains only the MLX-converted weights and tokenizer assets.
• Behavioral claims, evaluation numbers, and limitations are inherited from the source model.
• The source model is a model-editing research artifact with dual-use implications.
• The model inherits licensing considerations from the source and base model.

Source

• Source model: GestaltLabs/Ornstein-Hermes-3.6-27b-SABER
• Base model lineage: GestaltLabs/Ornstein-Hermes-3.6-27b