ENPT

Google Gemma points to djev on Cloud Run

Last updated:

Source · Model card: djev-spark

On 22 September 2026, @googlegemma posted that deploying DiffusionGemma-Jev (djev) got easier via taeold/djev-run: a single-command path to a Jev API–compatible endpoint on Google Cloud Run (NVIDIA RTX PRO 6000), built on mmastrac/djev-spark.

This is a deploy / hosting pack, not a new catalog model. ModelSystem.One already tracks the stack under djev-spark. The Cloud Run path lowers the “own a Spark box” barrier for experiments; it does not mint a separate Gemma System One entry on the quadrant.

UnverifiedClaim / hygiene: the tweet’s ~35–60 ms single-step latency, batch@32 ~100–123 rps, and ~$3/hr (idle → $0) figures are vendor/amplification numbers. The djev-run README also publishes JevBench v1.3.0 tables and cold-start timings — treat all as author-reported until re-run. Earlier @googlegemma “DiffusionGemma as Jev” coverage (18 Sep) remains the category-awareness piece.

Primary sources: X post · taeold/djev-run.