Nemotron 3.5 Bolt LiLiCorr draft, converted to a BF16 DFlash2 checkpoint

This is an unofficial conversion. It is not made, reviewed or endorsed by NVIDIA.

It's NVIDIA's released LiLiCorr draft for Nemotron 3.5 Bolt (74B-A9B, 24 GB), converted from the GGUF into the BF16 DFlash2 layout that Splash loads. I made it for the Nemotron 3.5 Bolt support proposed in incoai/splash#633, as an alternative to converting the GGUF at install time.

Source

Conversion

The weights were converted with install/dflash_gguf.py from the Splash branch ansoncat/splash@nemotron-3.5-bolt (commit f7211fe2; converter file sha256 69448bbe1fbbb5615e1fdc171cb14a97ccf3d15c22cd0c16f3f38f681363adef). The NVFP4, Q8_0, BF16 and F32 tensors are dequantized, scaled and rounded to BF16. No weights were retrained or changed in any other way.

model.safetensors here is byte-identical to what that branch's installer derives from the source GGUF:

  • model.safetensors: 2,288,586,972 bytes, sha256 05afa13e87a767b6c2f64246ba7c12fcde790dc5c76962f3a58dc961aef350b5

config.json is the DFlash2 config the installer generates from the GGUF header.

License

The source model is released under the OpenMDW License Agreement, version 1.1, and so is this conversion. A copy of the license is in LICENSE. The model and its original documentation are by NVIDIA; see the source repository for the model card.

Downloads last month
126
Safetensors
Model size
1B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ansoncat/Nemotron-3.5-Bolt-LiLiCorr-DFlash2