---

license: apache-2.0

base_model: JamieBradfield/qwen3.8-9b-hermes-fc-balanced

tags:

• qwen3.5
• function-calling
• tool-use
• rocmfp4
• gguf

pipeline_tag: text-generation

language:

• en

---

Qwen3.8-9B Hermes FC — Balanced (GGUF, ROCmFPX)

ROCmFPX-quantized GGUF of the balanced v26 fine-tune

(qwen3.8-9b-hermes-fc-balanced).

Files

• qwen3.8-9b-hf-fc-v26-Q4_0_ROCMFP4_FAST.gguf — Q4_0 with ROCmFP4 fast

kernels (AMD RDNA3). Target: ROCm llama.cpp inference on RX 7000-series GPUs.

How it was made

From the BF16 merge in the parent repo, using the llama-rocmfpx fork:

# 1. convert safetensors -> BF16 GGUF
python C:/AI/llama-rocmfpx/convert_hf_to_gguf.py <merged-dir> --outtype bf16

# 2. quantize BF16 -> Q4_0_ROCMFP4_FAST (ROCm 7.1 env; BF16 source needs no
#    --allow-requantize)
C:/AI/llama-rocmfpx/build-rdna3/bin/llama-quantize <src.gguf> <out.gguf> Q4_0_ROCMFP4_FAST

Served with `--ctx-size 245760 --n-gpu-layers 99 --flash-attn on

-ctk q8_0 -ctv turbo3`.

Notes

• ROCmFPX quants target AMD ROCm inference; for portable use, convert from

the BF16 merge in the parent repo.

• The merge preserves the 15-key MTP head; this quant carries it through.
• Model card, training recipe, and eval table: see the parent repo.