Add Qwythos-9B mixed-nvfp4 (Krill-native, qwen3_5)

a7aa5d3 verified about 10 hours ago

2.89 kB

license: apache-2.0
library_name: mlx
tags:
  - mlx
  - krill
  - qwen3_5
  - qwythos
  - vision-language
  - nvfp4
  - long-context
base_model: empero-ai/Qwythos-9B-Claude-Mythos-5-1M
pipeline_tag: image-text-to-text

Qwythos-9B-Claude-Mythos-5-1M — MLX nvfp4 (complete VLM, Krill-native)

A mixed-precision nvfp4 (group 16) quantization of empero-ai/Qwythos-9B-Claude-Mythos-5-1M, a Qwen3.5-class hybrid vision-language model.

Original model and weights by empero-ai (Qwythos-9B-Claude-Mythos-5-1M). Full credit to them; this repo only re-quantizes their model.

Why this build

👁️ Complete vision-language model — the vision tower is included. This build keeps the full VLM (text decoder + vision tower), not a text-only strip.
🎯 nvfp4 mixed precision. The decoder is nvfp4 at group size 16, with down_proj and o_proj protected at 8-bit and the vision tower kept at higher precision. Smaller and faster than int4 at comparable quality.
⚡ Native Krill runtime. Runs as a native Swift + MLX model on Apple Silicon, on Krill's from-scratch runtime for the Qwen3.5 hybrid GatedDeltaNet (SSM) + full-attention decoder — not an mlx_vlm passthrough.
🧵 Long context. 262K native (1M via YaRN rope-scaling upstream).

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 Qwythos nvfp4 (pulls this repo)
krill run qwythos-9b-nvfp4 "Give three tips for staying focused while studying."

krill update

Run with mlx_vlm (text + vision)

pip install -U mlx-vlm
python -m mlx_vlm generate --model srv-sngh/Qwythos-9B-Claude-Mythos-5-1M-mlx-nvfp4 \
  --prompt "Describe this image." --image path/to/image.jpg --max-tokens 200

About the base model

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

Quantization

field	value
format	MLX nvfp4 (mixed precision)
group size	16
protected	`down_proj`, `o_proj` @ 8-bit affine; vision tower at higher precision
size	~6.4 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).

License

apache-2.0, matching the base model empero-ai/Qwythos-9B-Claude-Mythos-5-1M.