Wondernutts/diffusiongemma-26B-A4B-it-openvino-int4

DiffusionGemma 26B-A4B, OpenVINO int4 (C=48 baseline)

The first OpenVINO port of a diffusion LLM (first text generated on an Intel Arc B70, 2026-07-10). Unified backbone IR: tied encoder/decoder weights deduped, capacity-dispatch MoE at C=48 slots per expert, int4 sym group 64, int8 backup, fp32 lm_head/softcap/self-conditioning tail.

This is the pristine pre-LUT export, kept as the full-capacity baseline. For positions past ~1K apply the RoPE sin/cos LUT graph patch (patches/dg_rope_lut.py).

Runs on Arc with default compile settings. Caveat: this build's quality numbers were measured with dynamic quantization on, which the C=32 successor showed to be damaging on this model (DYNAMIC_QUANTIZATION_GROUP_SIZE=0 produced sharper logits and fewer degenerate draws). DQ0 on this build is untested; try it first.

DG_UNIFIED_EXPORT.py here is the unified-export cell extracted from the full Colab notebook, kept as a build reference; the complete runnable recipe is colab/COLAB_DiffusionGemma.py. The sampler and server that run this artifact on Arc, plus running notes and the Arc driver bug catalog, are in the same repo (DIFFUSION_NOTES.md). The serving recipe and field-tested defaults also ship in sampler_manifest.json.

Faster successor with better coherence (C=32, requires DYNAMIC_QUANTIZATION_GROUP_SIZE=0): Wondernutts/diffusiongemma-26B-A4B-it-openvino-int4-c32

Toolkit and Arc bug catalog: OpenVino-For-Gemma-4