FP32 Casting: The original BF16 safetensors were cast to FP32 before conversion.
Architecture Patching: Used a patched version of llama.cpp (based on build b3962) with the city96 diffusion patch to support the cosmos architecture.
Precision Preservation: Sensitive layers (normalizations and biases) were kept in F32 to prevent color banding and generation artifacts.

💾 Available Quants

File Name	Quant Type	Size	Description
`anima-base-v1.0-Q8_0.gguf`	Q8_0	~2.1 GB	Near-perfect fidelity, recommended for 16GB+ RAM.
`anima-base-v1.0-Q6_K.gguf`	Q6_K	~1.6 GB	Excellent balance of speed and quality.
`anima-base-v1.0-Q5_K_M.gguf`	Q5_K_M	~1.4 GB	The "Sweet Spot" for most users.
`anima-base-v1.0-Q4_K_M.gguf`	Q4_K_M	~1.2 GB	Highly efficient, great for low-VRAM/RAM.
`anima-base-v1.0-Q3_K_M.gguf`	Q3_K_M	~1.0 GB	Maximum compression.

🧩 Required Components

To use these GGUF weights, you still need the following parts from the original Anima release:

This model follows the NVIDIA Cosmos License. Please refer to the original model card for full usage restrictions.