NVIDIA-Nemotron-3.5-Bolt-E4B — ROCmFPX GGUF

nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16, NVIDIA's small model post-trained from Google Gemma 4 E4B, quantized to the ROCmFPX formats, a llama.cpp fork with low-bit weight types for AMD GPUs (Vulkan and HIP/ROCm).

Start with Q6_0_ROCMFPX (6.2 GB). The 4-bit files are smaller but clearly further from the original.

These files need HarshaVardhan3002/ROCmFPX at commit b130d4995 or newer. Stock llama.cpp, LM Studio, Ollama and other ROCmFPX builds may not load them. These files have not been tested on AMD hardware yet.

Not an official NVIDIA release. All credit for the model goes to its authors.

Text only. The GGUFs here hold the language model. For images, use the vision projector (mmproj-*.gguf) from unsloth/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF with --mmproj (not tested with these files).

Files and measured quality

Reference: the BF16 model. Measured on WikiText-2 test, 40 chunks of 512 tokens. KLD is the mean KL divergence of the next-token distribution from BF16 (lower is better). Same top is how often the quant picks the same next token.

The 6-bit file is named NVIDIA-Nemotron-3.5-Bolt-E4B-Q6_K_ROCMFPX.gguf so that Hugging Face lists it as a 6-bit quant. Its weights are the ROCmFPX Q6_0_ROCMFPX type, not a stock llama.cpp Q6_K, and it needs the same fork as the other files.

For a small model like this, 4-bit costs a lot more quality than 6-bit, so use the 4-bit files only when memory is tight. All four files answer a simple chat question correctly.

Not published: Q3_0_ROCMFPX (4.21 GB, KLD 0.3582) is larger than Q4_0_ROCMFP4_FAST and further from BF16.

How these were made

1. BF16 GGUF converted with the fork's convert_hf_to_gguf.py.
2. Importance matrix from the BF16 file over a general calibration text (calibration_datav3), 512-token chunks.
3. llama-quantize --imatrix from the BF16 file, one run per type, from HarshaVardhan3002/ROCmFPX. No tensor scales were needed (the model's weights are large enough for the ROCmFPX block scales).
4. Quality: the BF16 reference ran on an NVIDIA GPU with stock kernels. Each quant was decoded by the fork's CPU code to F16 and scored on the same GPU, so the numbers measure the weights only. AMD GPU kernels compute the same math in a different order, so results on an AMD GPU can differ slightly.

Run

git clone https://github.com/HarshaVardhan3002/ROCmFPX && cd ROCmFPX
cmake -B build -DGGML_VULKAN=ON          # or -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1151 (gfx1100 for RX 7900)
cmake --build build --target llama-server -j
build/bin/llama-server -m NVIDIA-Nemotron-3.5-Bolt-E4B-Q6_K_ROCMFPX.gguf -ngl 999 -fa on --jinja -c 32768

License

OpenMDW-1.1, same as the base model. See the original model card for its terms and intended use.