Google Gemma amplifies “DiffusionGemma as Jev”
Source · Model card: djev-spark
On 18 September 2026, the official @googlegemma account posted “DiffusionGemma as Jev”, quoting @mmastrac’s live evals of DiffusionGemma-as-Jev via a vLLM structured-reads patch.
The post frames non-autoregressive canvas diffusion as a way to score structured choices in a single parallel pass (~0.2s on a DGX Spark in their wording), with bidirectional attention over options and multimodal grounding from the Gemma lineage. It is category marketing from Google’s open-model account, not a TypeSafe product launch.
On ModelSystem.One this maps to the existing catalog entry djev-spark — mmastrac’s DGX Spark / GB10 stack that serves DiffusionGemma behind a Jev-shaped POST /v1/systemone API (mmastrac/djev-spark). No separate “Gemma System One” model belongs on the quadrant.
UnverifiedClaim / hygiene: speed (~0.2s), “well-calibrated” distributions, and mmastrac’s earlier “DiffusionGemma wins / ties Jev” headlines mix hosted Jev API latency with local DGX Spark runs — different boundaries. Treat Google’s amplification as an adoption signal, not a ModelSystem.One quality bench.
Primary source: x.com/googlegemma/status/2101069861598482817. Upstream threads: mmastrac “Jev at home”, evals quote.