ZeroWw/granite-3.1-8b-instruct-GGUF

My own (ZeroWw) quantizations. output and embed tensors quantized to f16. all other tensors quantized to q5_k or q6_k.

Result: both f16.q6 and f16.q5 are smaller than q8_0 standard quantization and they perform as well as the pure f16.

Updated on: Sun Jan 05, 18:30:27