Granite SWASH 3B-a600M GGUF

GGUF conversion of:

ibm-granite/granite-swash-3b-a600m

Original model: https://huggingface.co/ibm-granite/granite-swash-3b-a600m

Architecture

• GraniteMoeSWAForCausalLM
• Mixture of Experts
• Sliding Window Attention
• Learnable Attention Sinks
• 3.02B total parameters
• ~600M active parameters per token
• 8192 context length

Files

BF16

granite-swash-3b-a600m-BF16.gguf

Converted using experimental GraniteMoeSWA support from llama.cpp PR #25505:

https://github.com/ggml-org/llama.cpp/pull/25505

License

Apache-2.0, following the original IBM Granite model.