ToPo-ToPo/Ornith-1.0-35B-mlx-4bit

MLX 4bit conversion of deepreinforce-ai/Ornith-1.0-35B for Apple Silicon (mlx-vlm).

Provenance (self-converted from official weights)

• Source: deepreinforce-ai/Ornith-1.0-35B (license: mit)
• Tool: mlx-vlm 0.6.3 — mlx_vlm.convert --hf-path deepreinforce-ai/Ornith-1.0-35B --mlx-path . -q --q-bits 4 --q-group-size 64
• Effective: 4.649 bits/weight
• Note (MoE): the source stores experts per-expert (experts.{i}.gate_proj/up_proj/down_proj), while mlx-vlm expects a fused/stacked experts.gate_up_proj. A sanitize monkeypatch fused them before conversion (gate then up, stacked over experts).
• Validation: reproduced geometrically exact CAD output in an agentic CAD+FEM pipeline (volumes match the reference mlx-community conversion).

Usage

from mlx_vlm import load, generate
model, processor = load("ToPo-ToPo/Ornith-1.0-35B-mlx-4bit")