mmastrac/diffgemma-26b-a4b-it-nvfp4x

DiffGemma 26B-A4B-it — nvfp4x (.dgq)

A quantized, self-contained .dgq pack of DiffusionGemma 26B-A4B-it (Gemma-4 26B-A4B MoE, discrete block diffusion) for the diffgemma-mps Rust + Metal inference engine on Apple Silicon. Text-only (v1).

nvfp4x = the q4 profile with the MoE experts pushed to NVFP4 instead of int4 affine (attention and the dense FFN stay bf16).

Quantization — nvfp4x

Tensor class	Format	Bits/weight
MoE experts	nvfp4_block (NVFP4, group 16)	~4.5
Attention (q/k/v/o), dense FFN	bf16	16
Embeddings, router, norms	bf16	16
Self-conditioning MLP	q8_row	8

~18 GiB of weights. Manifest version 4 (custom-class): requires a diffgemma-mps build with the --set class=format knob — older builds refuse it loudly rather than mis-dispatch.

Usage

diffgemma-mps download --repo mmastrac/diffgemma-26b-a4b-it-nvfp4x -o model/diffgemma-26b-a4b-it-nvfp4x
diffgemma-mps -m model/diffgemma-26b-a4b-it-nvfp4x chat

Build and run details: github.com/mmastrac/diffgemma.

License

Gemma Terms of Use. Derived from google/diffusiongemma-26B-A4B-it; use is subject to the Gemma license.