On June 10, 2026, Google released DiffusionGemma, an experimental open-weight model using a text-diffusion approach, under an Apache 2.0 license. Positioned as a speed-focused model, it aims for up to 4x faster generation through parallel token output.
June 10, 2026 · Google
DiffusionGemma generates text in parallel — up to 4× faster output
An experimental open-weight model that drops the one-token-at-a-time approach. Instead it generates and refines whole text blocks from noise, trading memory pressure for compute and adding self-correction mid-generation. Released under Apache 2.0.
26B
Total parameters — only ~3.8B active per inference (MoE)
24GB
Quantized version fits one consumer GPU for fast local inference
700+
Tokens/sec reported on an RTX 5090 (1000+ in some postings)
Generation speed on dedicated GPUs
Relative token output — each block = 1×
1×
Autoregressive
up to 4×
DiffusionGemma
How it differs from the standard model
DiffusionGemma
Text diffusion — parallel, non-autoregressive
Priority: speed
Best on compute-bound GPUs (e.g. NVIDIA)
Strong at code, math, self-correction
Gemma 4 (standard)
Autoregressive — one token at a time
Priority: quality
Best in memory-bound environments
General high-quality generation
The diffusion loop — refine, don't just predict
Start from noise (whole block)
→
Generate block in parallel
→
Revisit & self-correct mid-generation
Developers welcome it
Strong at code and math, practical thanks to self-correction, and a real step forward for fast local inference under the permissive Apache 2.0 license.
Google's own caveats
Quality may fall short of standard Gemma 4. It is experimental, benefits shrink in memory-bound setups, and detailed benchmarks remain limited.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…
