CoolFace
Modelpublic

JANGQ-AI/Ornith-1.0-35B-JANG_6M

sourceHugging Facemitupdated 19d agoView on Hugging Face
3likes111downloads
Model Card

<p align="center"> <img src="mlx-studio-logo.png" alt="MLX Studio" width="400"/> </p>

<p align="center"> <img src="jangq-logo.png" alt="JANGQ" width="200"/> </p>

<p align="center"> <a href="https://vmlx.net"><img src="vmlx-app.png" alt="vMLX — run JANG models on Apple Silicon" width="820"></a> </p>

<h3 align="center">⚡ All JANG models are meant to be run in <a href="https://vmlx.net">vMLX</a></h3>

<div align="center">

Ornith-1.0-35B · JANG_6M

Vision-language · Qwen3.5 hybrid backbone · 6-bit near-lossless mixed precision · ~26 GB

</div>

⚠️ Requires [MLX Studio](https://mlxstudio.com) (or the vMLX runtime) to run. Standard mlx_lm cannot load JANG bundles correctly — they store the Qwen3.5 RMSNorm un-shifted and rely on the runtime's +1 scale_shift + per-layer bit detection. MLX Studio includes the JANG loader.

JANG_6M = 8-bit attention + 6-bit routed experts (affine mixed precision, group-size 64); vision tower kept fp16. A near-lossless JANG profile from JANGQ-AI.

Architecture

Familyqwen3_5_moe (hybrid)
Text layers40 — 30 Gated-DeltaNet + 10 full-attention
MoE / dims256 routed experts (stacked switch_mlp) · hidden 2048
VisionViT tower (model.visual) preserved fp16
Cachehybrid (GDN state + KV for attention layers)
Parsersreasoning qwen3 · tools qwen

Provenance

  • —Base: deepreinforce-ai/Ornith-1.0-35B © DeepReinforce — MIT (Qwen3.5-based)
  • —Quantization: JANG · JANG_6M (8-bit attention + 6-bit routed experts, group-size 64; vision tower fp16) · eric@jangq.ai