Breaking

Google Releases DiffusionGemma, a Parallel Text-Generation Model

June 10, 2026 at 14:06 EDT

Foundation Models
Open Source
Infra & Chips

Google DeepMind on June 10, 2026 released DiffusionGemma (google/diffusiongemma-26B-A4B-it), an experimental open model that generates text in parallel, under the Apache 2.0 license. NVIDIA announced day-0 support the same day.

June 10, 2026 · Google DeepMind

DiffusionGemma: Text Generation Goes Parallel

An open-weight (Apache 2.0) experimental model that generates text from noise in a single pass — producing 256 tokens in parallel per step instead of one token at a time. NVIDIA delivered same-day optimization for H100, DGX Spark, and RTX 5090.

256

tokens generated in parallel per step

26B / 3.8B

total / active params (MoE)

256K

max context length (tokens)

18–24GB

VRAM needed via quantization

Inference speed by platform

tokens per second (TPS) — taller column = faster

1,000+

H100

700+

RTX 5090

150+

DGX Spark

Autoregressive vs. Diffusion

Conventional LLM

one token at a time → memory-bandwidth bottleneck

DiffusionGemma

256 tokens per step from noise → high throughput

Where it shines

Code infilling & inline editing
Summaries and headline generation
Runs locally on DGX Spark / RTX — no cloud

The caveats

Output quality may trail standard Gemma 4
Peak speed only on compute-bound hardware
Speed-oriented, experimental release

Day-one delivery: BF16 + NVFP4 checkpoints · vLLM (FP8) support · free GPU endpoint · Apache 2.0 license

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.