back to local model checker

DiffusionGemma 26B-A4B IT

MoEToolsThinkingVision

Google · 26B (4B active) · Mixture of Experts

Discrete diffusion MoE — 1100+ tok/s on H100, multimodal (text/image/video)

Apache 2.0
Downloads

0

Likes

0

Released

2026-06

Context

256K tokens

Use cases

chatvisionreasoning
Total experts128
Active experts8
Active params3.8B

Your hardware

16 GB Reference Desktop

Quant
Q4_K_M
Memory
13.8 GB
Speed
~15 tok/s
Fit
Barely runs

Quantization options

What is quantization?
QuantBitsVRAMDiskQualityStatus
Q2_K28.8 GB7.9 GBlowDecent · ~23 t/s
Q3_K_M312.2 GB11.1 GBmoderateTight fit · ~17 t/s
Q4_K_M413.8 GB12.7 GBgoodBarely runs · ~15 t/s
Q5_K_M517.1 GB15.9 GBgoodToo heavy · ~8 t/s
Q6_K620.5 GB19.1 GBexcellentBarely runs · ~4 t/s
Q8_0827.1 GB25.4 GBexcellentBarely runs · ~2 t/s
F161653.8 GB50.9 GBlosslessToo heavy · ~1 t/s

About this model

DiffusionGemma 26B-A4B IT (DiffusionGemma) is a 26B sparse mixture-of-experts (MoE) model activating 4B active per token from Google in the Gemma family, designed primarily for chat, vision, reasoning workflows. Native reasoning/thinking tokens expand output length per query, so reserve extra KV cache headroom. Supports structured tool calling; use Q4_K_M or higher when strict JSON schema adherence matters. Supports up to 256K context tokens; start at 4K–8K with Q8_0 KV cache on tight GPUs.

Hardware planning baselines start at 14.5 GB minimum memory and 24.2 GB recommended memory. At Q4_K_M, weights plus baseline runtime overhead use roughly 13.8 GB of RAM/VRAM and 12.7 GB of disk space. Because MoE routing evaluates only a subset of experts on each generated token, DiffusionGemma 26B-A4B IT delivers faster decoding throughput per gigabyte of memory bandwidth than a dense checkpoint of identical total size—provided the full 13.8 GB checkpoint stays resident without SSD paging.

Deploy DiffusionGemma 26B-A4B IT locally via llama.cpp with Flash Attention enabled. Under Apache License 2.0, review official weight cards before commercial redistribution.

Highlights

  • Local Fit Baseline14.5 GB minimum and 24.2 GB recommended RAM/VRAM.
  • Primary QuantQ4_K_M targets 13.8 GB memory and 12.7 GB disk.
  • Context & Capabilities256K max context · Sparse MoE · Tool calling · Reasoning.

System requirements by device

Memory planning targets 14.5 GB minimum RAM/VRAM and 24.2 GB recommended memory. The baseline quant is Q4_K_M, estimated at 13.8 GB memory and 12.7 GB disk. Actual runtime usage shifts with context length, KV cache precision, and OS background overhead.

DeviceNotes
Mac (Apple silicon)Use unified memory headroom (~75% physical RAM max) as the primary fit signal. Start with Q4_K_M and close heavy browser tabs first.
Windows / Linux GPUKeep weights plus KV cache inside dedicated VRAM so Windows WDDM does not spill layers over PCIe into Shared GPU Memory.
Phones & Compact Edge BoardsNot practical on phones (14.5 GB min RAM exceeds Android LMK and iOS Jetsam limits). Pick a 2B–4B model on mobile.

Start with a short session and leave memory for the operating system. A file fitting on disk does not guarantee that it will load in memory.

Estimated memory by quantization

These are rough estimates, not tested minimum requirements. Check the actual file and runtime documentation; context, offloading and operating-system overhead can change the memory needed.

Quant TierEstimated memoryDisk SizeQuality
Q2_K8.8 GB7.9 GBlow
Q3_K_M12.2 GB11.1 GBmoderate
Q4_K_M13.8 GB12.7 GBgood
Q5_K_M17.1 GB15.9 GBgood
Q6_K20.5 GB19.1 GBexcellent
Q8_027.1 GB25.4 GBexcellent
F1653.8 GB50.9 GBlossless

How to run DiffusionGemma 26B-A4B IT on your device

Start with a short prompt and the smallest quant that still gives the quality you need. A clean first run is more useful than loading a larger file with no memory headroom.

Mac

Run DiffusionGemma 26B-A4B IT on Mac

  1. Install LM Studio or llama.cpp on macOS Apple Silicon.
  2. Pull Q4_K_M (13.8 GB) first and confirm it stays below ~75% of your Mac's unified memory.
  3. Enable Flash Attention and cap initial context at 8K until memory pressure stays below 24.2 GB.
  4. Consult the Mac guide for Activity Monitor tuning and Metal offload tips.
Device how-to guide
Windows

Run DiffusionGemma 26B-A4B IT on Windows

  1. Install LM Studio for Windows for visual VRAM inspection, or Ollama when you want a localhost API endpoint.
  2. Load Q4_K_M (~13.8 GB) and verify Dedicated GPU Memory stays within VRAM without spilling into Shared GPU Memory.
  3. Reserve at least 14.5 GB memory minimum and target 24.2 GB for comfortable context growth.
  4. Enable Q8_0 KV cache or drop one quant tier if generation slows down over PCIe.
Device how-to guide

DiffusionGemma 26B-A4B IT vs alternatives

ModelParamsContextMin RAMLicense
DiffusionGemma 26B-A4B IT26B (4B active)256K14.5 GBApache 2.0
Gemma 2 27B27B8K15.1 GBGemma
Gemma 3 27B27B128K15.1 GBGemma
Gemma 4 26B-A4B IT27B (4B active)256K15.1 GBGemma
Gemma 4 26B-A4B27B (4B active)256K15.1 GBGemma
Gemma 4 31B IT33B256K18.4 GBGemma

Use this table to choose the smallest nearby model that still matches the job. A related model with lower RAM can be the better local choice when the current page fits only at the edge of available memory.

Frequently asked questions

What is DiffusionGemma 26B-A4B IT?
DiffusionGemma 26B-A4B IT is a 26B sparse mixture-of-experts model (4B active active per token) from Google, supporting up to 256K tokens with 14.5 GB minimum RAM and 24.2 GB recommended RAM.
How do I download and install DiffusionGemma?
Download a GGUF build for DiffusionGemma 26B-A4B IT, start with Q4_K_M, and confirm the model fits locally before increasing context or quant quality.
How much VRAM or RAM does DiffusionGemma 26B-A4B IT require?
At Q4_K_M, DiffusionGemma 26B-A4B IT occupies roughly 13.8 GB of active memory and 12.7 GB on disk. Add 2–6 GB of spare headroom for KV cache buffers, runtime scratch space, and OS apps (24.2 GB recommended).
What is the best way to run DiffusionGemma 26B-A4B IT locally?
Run DiffusionGemma 26B-A4B IT in llama.cpp starting with Q4_K_M. Keep initial context at 4K–8K tokens and confirm 100% GPU/Metal offload before scaling context length.
Can DiffusionGemma 26B-A4B IT run on Android or iPhone?
DiffusionGemma 26B-A4B IT requires 14.5 GB minimum memory, placing it outside practical 8 GB smartphone limits. Run it on a 16 GB+ Mac or desktop GPU instead, and pick a 2B–4B checkpoint for phones.
How does DiffusionGemma 26B-A4B IT compare to Gemma 2 27B?
Gemma 2 27B (27B, 15.1 GB min RAM) sits closest in size. Compare active parameters, context ceiling (256K vs 8K), and license terms (Apache 2.0) before switching.