srv-sngh/Ornith-1.0-9B-MLX-4bit

Ornith-1.0-9B - MLX int4 (complete VLM, Krill-native)

A 4-bit (MLX affine, group 64) quantization of deepreinforce-ai/Ornith-1.0-9B - a Qwen3.5-class hybrid vision-language model.

Why this build

• 👁️ Complete vision-language model - the vision tower is included. Most community MLX/4-bit Ornith quants are text-only (vision stripped). This one keeps the full VLM, so it does image + text, not just text.
• ⚡ Native Krill runtime (from Krill v0.14). Runs as a native Swift + MLX model on Apple Silicon via krill run ornith-9b - Krill ships a from-scratch native runtime for Ornith's hybrid GatedDeltaNet (SSM) + attention decoder, not just an mlx_vlm passthrough.
• ✅ Parity-verified. Token-for-token match against mlx_vlm on the reference checkpoint.
• 💻 Fits a 24 GB Apple-silicon box at ~6 GB; one-line install + self-update (krill update).

Run in Krill (recommended)

# install Krill
brew tap srvsngh99/krill && brew install krill
# or:
curl -fsSL https://raw.githubusercontent.com/srvsngh99/Krill/main/install.sh | sh

# run Ornith (pulls this repo)
krill run ornith-9b "Give three tips for staying focused while studying."

# keep Krill up to date
krill update

Run with mlx_vlm (text + vision)

pip install -U mlx-vlm
python -m mlx_vlm generate --model srv-sngh/Ornith-1.0-9B-MLX-4bit \
  --prompt "Describe this image." --image path/to/image.jpg --max-tokens 200

About Ornith-1.0-9B

A Qwen3.5-class hybrid VLM: the text decoder is a Qwen3-Next-style stack interleaving GatedDeltaNet linear-attention (SSM) layers with full softmax-attention every fourth layer, plus a vision tower. Full credit to the original creators, deepreinforce-ai.

Quantization

field	value
format	MLX (affine)
bits	4
group size	64
size	~6 GB
contents	complete VLM (text decoder + vision tower)

In Krill, the text decoder runs natively; the vision tower currently runs via mlx_vlm (native vision is a follow-up). An nvfp4 build follows in Krill v0.14.1.

License

MIT, matching the base model deepreinforce-ai/Ornith-1.0-9B.