The source model contains experimental H-Res tensors named model.layers.*.hres.*. Current GGUF / llama.cpp runtimes do not implement the nonlinear H-Res residual adapter path, so these GGUF files omit the unsupported H-Res tensors during conversion.

For full H-Res behavior, use the original BF16 safetensors model: UraionLabs/Uraion-Agent-Steer.

Quantizations

Quant	File	Size	Suggested use
Q2_K	`Uraion-Agent-Steer-Q2_K.gguf`	2.81 GiB	Smallest fallback, lowest quality
Q3_K_M	`Uraion-Agent-Steer-Q3_K_M.gguf`	3.55 GiB	Low-memory fallback
Q4_K_M	`Uraion-Agent-Steer-Q4_K_M.gguf`	4.36 GiB	Recommended default for 8 GB GPUs
Q5_K_M	`Uraion-Agent-Steer-Q5_K_M.gguf`	5.07 GiB	Better quality if memory allows
Q6_K	`Uraion-Agent-Steer-Q6_K.gguf`	5.82 GiB	High quality local inference
Q8_0	`Uraion-Agent-Steer-Q8_0.gguf`	7.54 GiB	Largest compatibility quant

llama.cpp

bash

llama-cli -hf UraionLabs/Uraion-Agent-Steer-GGUF:Q4_K_M \
  -p "What is tool calling?"

Ollama

ollama run hf.co/UraionLabs/Uraion-Agent-Steer-GGUF:Q4_K_M

Conversion

These files were produced from a BF16 GGUF intermediate using llama.cpp quantization types: Q2_K, Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, and Q8_0.