Ornith 1.0 35B — Dynamic NeuroPack (NPK)
The 35B MoE Ornith coder packed with the NeuroPack dynamic codec for the Sentinel Coder One VS Code extension. Dynamic per-tensor quantization (Vulkan-safe palette):
- Attention (q/k/v/o), token embeddings and output head kept at Q8_0 (8-bit, quality-critical)
- FFN / MoE expert bulk at Q4_0 (4-bit) — where the size lives
- Norms at F32
Size
| Build | Size | vs bf16 (69.4 GB) |
|---|---|---|
ornith-35b-dynamic.npk |
20.48 GB | ~70% smaller |
Runs GPU-only on a ≥24 GB GPU (RTX 4090/3090/A5000+). It does not fit a
12 GB card GPU-only (the Vulkan-safe 4-bit floor for 35B is ~20 GB). For 12 GB GPUs
use the 9B dynamic NPK (qubitpage/ornith-9b-classic-gguf, ~5.3 GB).
Use
In Sentinel Coder: Add NeuroPack (.npk) Model from Folder → select the folder. Full agentic / tools / RAG, no Ollama, no server.
Converter + dynamic codec: https://github.com/msrusu87/npk-converter
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support