DiffusionGemma denoises in a single parallel pass, ~0.2s structured decisions on DGX Spark
inductionheads · x · 2026-09-20
The official Google Gemma account introduces "DiffusionGemma as Jev," a demo showcasing non-autoregressive architectures: instead of sequential token generation, canvas diffusion evaluates structured choices in a single parallel pass.
- Massive parallelism: denoises across an open canvas in one step, 0.2s on a DGX Spark
- Full bidirectional attention: every option attends to the full context concurrently, yielding well-calibrated decision distributions
- Multimodal grounding: inherits Gemma 4's spatial vision for complex visual and text decisions
Google also asks if users want DiffusionGemma on the Gemini API.
Related event: Google's DiffusionGemma Makes Decisions in 0.2s via Parallel Denoising(2 posts)→
More from Models
- Blogger vouches for rumored Gemini 4 Carbon: good taste, cost-effective — intellectronica · 2026-11-12
- Walden Robotics' Sage model quadruples robot uptime on Toyota factory lines — adnothing · 2026-11-12
- Gemini won't let you keep chat history while opting out of AI training — Google bundles both into one setting — arshdeep · 2026-11-12
- Gemini 4 Carbon Flex tier now costs the same per token as GPT 6.1 Terra — mariofilhoml · 2026-11-12
- Real-world test: Kimi K3.1 Pro costs 5.4x less than Opus 5.5 per task despite 2x list-price gap — Arindam_1729 · 2026-11-12
- Claude Fable 5.2 tops Agent Arena at +30.4%, 1.5x the gain of runner-up Opus 5.5 — daniel_mac8 · 2026-11-12