Cosmos-Reason2-8B-GGUF

GGUF conversion of nvidia/Cosmos-Reason2-8B.

Files

• Cosmos-Reason2-8B-BF16-split-00001-of-00005.gguf to ...-00005-of-00005.gguf: Main model (BF16, split into 5 parts)
• mmproj-Cosmos-Reason2-8B-BF16.gguf: Vision encoder (multimodal projector)

Download

huggingface-cli download robertzty/Cosmos-Reason2-8B-GGUF --local-dir ./Cosmos-Reason2-8B-GGUF

Usage with llama.cpp

No need to merge split files - llama.cpp loads them automatically:

llama-cli -m ./Cosmos-Reason2-8B-GGUF/Cosmos-Reason2-8B-BF16-split-00001-of-00005.gguf \
  --mmproj ./Cosmos-Reason2-8B-GGUF/mmproj-Cosmos-Reason2-8B-BF16.gguf -cnv

Optional: merge into single file:

llama-gguf-split --merge Cosmos-Reason2-8B-BF16-split-00001-of-00005.gguf Cosmos-Reason2-8B-BF16.gguf

Notes

• Requires llama.cpp b7480+ for qwen3vl architecture support
• Ollama and LM Studio may have compatibility issues with qwen3vl models