ai-trend-notifier
← wiki

$ cat wiki/models/diffusion-gemma.md

DiffusionGemma

Google DeepMind's first open-weight text diffusion model, built on the Gemma 4 26B MoE backbone.

Spec

AttributeValue
DeveloperGoogle DeepMind
Released2026-06-11
Announced2026-06-11
Context windowunknown
Pricingfree (open-weight)
LicenseApache 2.0
AvailabilityHugging Face (google/diffusiongemma-26B-A4B-it), Kaggle, Google Cloud Vertex AI Model Garden

Release Date

June 11, 2026.

Architecture

  • Base: Gemma 4 26B-A4B Mixture-of-Experts (25.2B total params, ~3.8B active per forward pass)
  • Decoding: Discrete text diffusion — generates 256-token blocks in parallel via iterative denoising (Uniform State Diffusion), not left-to-right autoregressive
  • Memory: ~18 GB VRAM (MXFP4 quantization) — significantly lower local deployment barrier than comparable models
  • Framework: native vLLM support

Benchmarks

DiffusionGemma trades accuracy for throughput vs. the autoregressive Gemma 4 26B-A4B baseline:

BenchmarkDiffusionGemmaGemma 4 26B-A4B (autoregressive)
AIME 202669.1%88.3%
MMMU Pro54.3%73.8%
Multi-step reasoning (math, vision-language) takes a larger accuracy hit. Pure text tasks suffer less. The tradeoff is the ~4× throughput gain.

Performance

  • 1,000+ tokens/sec on a single H100 GPU
  • 700+ tokens/sec on RTX 5090
  • ~4× faster than autoregressive Gemma 4 26B-A4B at matched benchmark quality for throughput-bound workloads

Use Cases

  • High-throughput text generation workloads where latency or cost per token is the primary constraint
  • Not suited for tasks requiring maximum reasoning accuracy (multi-step math, complex vision-language)

Compared To

ModelTypeSpeed (H100)AIME 2026License
DiffusionGemmaDiffusion MoE1,000+ t/s69.1%Apache 2.0
Gemma 4 26B-A4BAutoregressive MoE~250 t/s (est.)88.3%Apache 2.0
Kimi K3Autoregressive MoEunknownModified MIT

Significance

DiffusionGemma is the first open-weight text diffusion model from a major AI lab to enter production. Parallel block-decoding (256 tokens per iteration) challenges autoregressive as the only viable decode path at scale. Opens engineering trade-space: operators can now choose throughput over per-task reasoning quality, a tradeoff that didn't exist in open-weight production models before this release.

Referenced by

Sources