DiffusionGemma 26B-A4B

Name: DiffusionGemma 26B-A4B
Author: Google

Google · Gemma 4 · released 2026-06-09 · data checked 2026-11-01

DiffusionGemma 26B-A4B is a 25.8B-parameter Mixture-of-Experts model (4B active per token) from Google (United States), released 2026-06-09 under the Apache-2.0 license. It scores 52.6 on the Small LM Index, #25 of 34 medium models. It needs about 17.2 GB of memory at 4-bit with an 8K-token context.

Diffusion language model: writes blocks of text in parallel instead of one token at a time, trading some accuracy for speed.

Small LM Index

52.6

#25 of 34 scored medium models

Likely 48.8–56.3 · from 5 of 6 core benchmarks

Parameters

25.8B

4B active · MoE

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

256K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best medium model on that benchmark. Sources: 8 reported by Google. How we label sources.

MMLU-Pro Lab-reported Knowledge77.6

Class best: Qwen3.6 27B 86.2

GPQA Diamond Lab-reported Knowledge73.2

Class best: Nanbeige4.3 33B 89.5

Humanity's Last Exam Lab-reported Reasoning11.0

Class best: Qwen3.8 27B 30.8

AIME Lab-reported Math · AIME 2026, no tools69.1

Class best: Ternary Bonsai 2 27B 95.8

LiveCodeBench Lab-reported Coding · LiveCodeBench v669.1

Class best: Qwen3.8 27B 90.3

τ²-Bench Lab-reported Agentic · Average over 3 domains56.2

Class best: MiniCPM-V 4.7 35B-A3B 84.7

MMMU-Pro Lab-reported Vision54.3

Class best: Gemma 4 31B 76.9

Multilingual MMLU Lab-reported Multilingual · MMMLU81.5

Class best: EXAONE 4.0 32B 85.6

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

SWE-bench Verified≈ 35

Specifications

Lab: Google (US · United States)
Architecture: Mixture-of-Experts
Total parameters: 25.8B
Active parameters: 4B
Size class: Medium
Context window: 256K tokens
Input modalities: text, image
Reasoning: Hybrid (thinking on/off)
License: Apache-2.0
Openness: Open weights
Memory to run: ~17.2 GB at 4-bit, 8K context (more sizes)
Hugging Face: 625.6K downloads (30 days) · 1.3K likes

Weights on Hugging Face ↗

Compare with

DiffusionGemma 26B-A4B vs Gemma 4 26B-A4B (25.2B, index 67.5)
DiffusionGemma 26B-A4B vs Qwen3.8 27B (27B, index 77.0)
DiffusionGemma 26B-A4B vs Qwen3.6 27B (27B, index 74.0)
DiffusionGemma 26B-A4B vs Qwen3.5 27B (27B, index 70.6)

Sources

https://huggingface.co/google/diffusiongemma-26B-A4B-it

Something wrong here? Report it.

Run it locally

Uses unsloth/diffusiongemma-26B-A4B-it-GGUF (Q4_K_M). Swap the tag for another size: BF16, Q4_K_M, Q5_K_M, Q6_K, Q8_0.

Ollama

ollama run hf.co/unsloth/diffusiongemma-26B-A4B-it-GGUF:Q4_K_M

llama.cpp

llama-server -hf unsloth/diffusiongemma-26B-A4B-it-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/diffusiongemma-26B-A4B-it-4bit

Text only; for images use mlx-vlm.

vLLM (GPU server)

vllm serve google/diffusiongemma-26B-A4B-it

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	256K context
16-bit F16	53 GB	54 GB	63 GB
8-bit Q8_0	29 GB	30 GB	39 GB
4-bit Q4_K_M	17 GB	18 GB	27 GB

Check it against your GPU, Mac or phone →