Ollama v0.19.0, macOS Apple Silicon

cat > Modelfile <<EOF2 FROM /path/to/Nemotron-3-Super-120B-A12B-MLX-6bit

TEMPLATE """{{- if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}<|im_start|>user {{ .Prompt }}<|im_end|> <|im_start|>assistant """

PARAMETER stop "<|im_end|>" PARAMETER stop "</s>" PARAMETER num_ctx 8192 EOF2

ollama create nemotron-120b -f Modelfile

Error: unknown data type: U32

TL;DR

The most likely fix is to add support for U32 data type in the ollama create command by utilizing the tensorImportTransform framework.

Guidance

• The error message Error: unknown data type: U32 suggests that the ollama create command does not currently support the U32 data type used in the MLX-quantized SafeTensors for Nemotron-H models.
• To fix this, the tensorImportTransform framework from PR #14878 can be used to add support for U32 data type, similar to how it was done for Qwen3.5 support.
• The architecture class for the registry should be NemotronHForCausalLM to enable native MLX import for Nemotron-H models.
• Verifying the fix can be done by running the ollama create command with the updated tensorImportTransform framework and checking if the error message is resolved.

Example

No code snippet is provided as it is not clearly supported by the issue, but the tensorImportTransform framework can be used as a reference to add support for U32 data type.

Notes

The current limitation is that the ollama create command does not support U32 data type, which is used in MLX-quantized SafeTensors for Nemotron-H models. Adding support for U32 data type using the tensorImportTransform framework should resolve the issue.

Recommendation

Apply workaround by utilizing the tensorImportTransform framework to add support for U32 data type, as it is a straightforward solution that has been successfully implemented for Qwen3.5 support.