A Jev-compatible API for text, images and video

Prem offers a zero-data-retention decision API powered by Google’s experimental open DiffusionGemma model, using TypeSafe’s Jev-compatible SystemOne interface to make fast, bounded workflow judgments from text, images, or short videos. Applications can submit up to 16 independent questions per request and receive structured Choice, Noul yes/no probability, or Score rubric responses, enabling routing, escalation, document review, support triage, and other rule-based actions while retaining human oversight. Prem manages authentication, serving, rate limits, accounting, and multimodal input preparation, while requests are processed in plaintext without confidential enclave execution and neither Prem nor its serving partner retains prompts or completions. The service supports a single image or an MP4 subject to size and duration limits, samples up to four video frames, and does not process audio, while users are advised to validate model behavior because probabilities are normalized scores rather than calibrated accuracy estimates and can be affected by option order. Available through a playground and API, the deployment runs on one 96 GB Blackwell GPU, achieved 83.5% accuracy on 231 public JevBench items, and recorded roughly 13 requests per second with 440 ms p95 latency in a local test, though the company emphasizes that customers should evaluate accuracy and action thresholds on representative workloads.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.