willitrun·ai

GoogleGoogle

DiffusionGemma 26B A4B

Frontier
huggingfaceHuggingFace
1.7MDownloads1.1KLikesJun 2026Veröffentlicht262K TokenKontextApache 2.0Lizenz68 GutQualität

DiffusionGemma 26B A4B (25.799999237060547B parameters) requires approximately 22.4 GB of VRAM with Q4_K_M quantization. As a Mixture of Experts model with 4B active parameters, it uses less memory than its total parameter count suggests. For the best balance of quality and speed, we recommend hardware with at least 26 GB of VRAM.

Loslegen

— kopieren & einfügen, um lokal auszuführen

Copy-paste commands to run DiffusionGemma 26B A4B on your machine.

Run

docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \ --hf-repo "google/diffusiongemma-26B-A4B-it" \ --hf-file "diffusiongemma-26B-A4B-it-Q4_K_M.gguf" \ -c 4096 -ngl 99

Quick specs

Parameters25.8B (4B active)
Architecturemoe (MoE)
Context262K tokens
Modalitytext
Min RAM10.1 GB
Rec. RAM15.7 GB (Q4_K_M)
LicenseApache 2.0
FamilyGemma
✓ Vision✓ Chat✓ Reasoning

About this model

DiffusionGemma 26B A4B is Google's block-diffusion language model in the Gemma family: instead of left-to-right autoregression it denoises blocks of tokens in parallel. 25.8B total parameters with ~4B activated per token (128 experts, 8 active) and multimodal image-text input.

  • •Block-diffusion decoding — generates blocks of tokens in parallel rather than strictly left-to-right.
  • •MoE efficiency: 128 experts, 8 active per token (~4B activated of 26B total).
  • •Gemma-based multimodal (image-text) backbone with 256K context.
  • •Research-oriented alternative to autoregressive LLMs.

Verwandte Modelle

Deine Hardware

Erkennung...

Schnellauswahl

Bestes BudgetA
Mac mini M4 64GB~$1,099 — 10 tok/s
NVIDIA
Beste GesamtwahlS
RTX 5090 32GB~$1,999 — 144 tok/s

Beste Hardware

Top-Empfehlungen für DiffusionGemma 26B A4B

NVIDIA
RTX 5090 32GBS
32 GB
AMD Instinct MI100 32GBS
32 GB
NVIDIA
RTX PRO 4500 Blackwell 32GBS
32 GB
NVIDIA
NVIDIA V100 32GBA
32 GB
NVIDIA
NVIDIA A100 40GBA
40 GB

Dieses Modell ausführen

DiffusionGemma 26B A4B on RTX 5090 32GBDiffusionGemma 26B A4B on AMD Instinct MI100 32GBDiffusionGemma 26B A4B on RTX PRO 4500 Blackwell 32GB

Quantisierungsoptionen

VRAM-Schätzungen nach Quantisierungsstufe

No hardware detected — fit column shows raw VRAM estimates

QuantBitsVRAMQualityFit
Q2_K
2
10.1 GB
Low—
Q3_K_S
3
12.6 GB
Low—
NVFP4
4
14.4 GB
Medium—
Q4_K_M
4
15.7 GB
Medium—
Q5_K_M
5
18.6 GB
High—
Q6_K
6
21.2 GB
High—
Q8_0
8
27.6 GB
Very High—
F16
16
52.9 GB
Maximum—

Hardware-Kompatibilität

Eignungsschätzungen für alle Hardware

Rechner öffnen

Computing compatibility...

Speicheraufschlüsselung

Reference: RTX 2060 6GB

Weights15.7 GB
KV Cache3.7 GB
Runtime2.4 GB
Headroom0.6 GB

Häufig gestellte Fragen

FAQ — DiffusionGemma 26B A4B

Siehe auch

QuantisierungsleitfadenBewertungsmethodikRechner öffnen