diffusiongemma-26B-A4B-it

Home/Models/google/diffusiongemma-26B-A4B-it

diffusiongemma-26B-A4B-it

google/diffusiongemma-26B-A4B-it

kind: original
quant: bf16 (16-bit)
family: gemma
modality: image-text-to-text
root model: google/diffusiongemma-26B-A4B-it
provider: google
license: apache-2.0
language: —
context: —
library: transformers
vRAM (GB): —
size on disk: —
downloads: 1,871,257
likes: 1,147
created: —
last modified: 2026-07-15T16:35:45+00:00 (2026)
huggingface: https://huggingface.co/google/diffusiongemma-26B-A4B-it
tags: transformerssafetensorsdiffusion_gemmaimage-text-to-textconversationallicense:apache-2.0endpoints_compatibleregion:us

Benchmark scores (9)

benchmark	category	score	context	conf	evidence
MMLU-Pro	knowledge	77.6	MMLU Pro	medium	`\| MMLU Pro \| 77.6% \| 82.6% \|`
GPQA Diamond	reasoning	73.2	GPQA Diamond	medium	`\| GPQA Diamond \| 73.2% \| 82.3% \|`
MATH	math	70.5	MATH-Vision	medium	`\| MATH-Vision \| 70.5% \| 82.4% \|`
AIME 2026	math	69.1	AIME 2026 no tools	medium	`\| AIME 2026 no tools \| 69.1% \| 88.3% \|`
LiveCodeBench	coding	69.1	LiveCodeBench v6	medium	`\| LiveCodeBench v6 \| 69.1% \| 77.1% \|`
MMMU	multimodal	54.3	MMMU Pro	medium	`\| MMMU Pro \| 54.3% \| 73.8% \|`
BigBench	reasoning	47.6	BigBench Extra Hard	medium	`\| BigBench Extra Hard \| 47.6% \| 64.8% \|`
Humanity's Last Exam	knowledge	11.9	HLE with search	medium	`\| HLE with search \| 11.9% \| 17.2% \|`
Humanity's Last Exam	knowledge	11.0	HLE no tools	medium	`\| HLE no tools \| 11.0% \| 8.7% \|`

Discussions (23)

title	author	status	type
Raise per-image vision soft-token budget from 280 to 1120	lucianommartins	open	disc
fix: chat template — null handling, reasoning preservation, turn-tag balance, input validation	joaogante	merged	disc
License clarification	Lounafa1	closed	disc
Field report: DiffusionGemma 26B as a full Claude Code agent on a single RTX 5090 (WSL2) — what worked, and the guardrails for the accuracy problem	Markbdean	open	disc
Add files for diffusers loading	dg845	merged	disc
Batch Invariance	muki987	open	disc
Bien	Ixarx	open	disc
Architectural proposal: Dynamic "Endocrine" Neuromodulation & Homeostatic Flushing for DiffusionGemma	Soulfate24	open	disc
Custom chat template: empty content in multi-turn conversations (with fix)	buzman	open	disc
Update non-thinking chat template	joaogante	merged	disc
Problem with tensor parallel size==2	borodulinaad	open	disc
DiffusionGemma	razielAI	open	disc
Progressive Hierarchical Canvas Diffusion	neoxider2	open	disc
VLLM setup? Getting malformed responses. What is correct configuration	jtvino	open	disc
Update README.md	cantap555	open	disc
ADD FAST-dLLM mode for more speed	VaLtEc-BoY	open	disc
worlds most quantization sensitive model, congrats🥳🤩	unokayish182	open	disc
PLEASE E4B	VaLtEc-BoY	open	disc
RuntimeError: Tensor.item() cannot be called on meta tensors	kimyong95	open	disc
Argument input_ids not found in the forward method	mohamedemam	open	disc
Runtime error and Module not found error	harishreddy09876543	open	disc
Stronger model for Agentic workflow	Duonglv	open	disc
ModuleNotFoundError: Could not import module 'DiffusionGemmaForBlockDiffusion'. Are this object's requirements defined correctly?	Phase-Technologies	open	disc