Celeris-1 Decision Hits 80% on JevBench
Celeris-1 Decision is a multimodal diffusion model for bounded, structured decisions over text, JSON, and images. Celeris says it reaches 80% accuracy on JevBench—7 points above Jev 1.13.0—with 67ms median latency.
Celeris is targeting a compelling gap between classifiers and general-purpose LLMs: fast, typed decisions that application code can consume directly.
- –System One API compatibility lets existing Jev integrations switch with minimal code changes
- –Outputs probabilities, choices, and scores instead of free-form text, reducing parsing and routing overhead
- –Multimodal inputs expand use cases to receipt checks, document review, visual verification, and agent control
- –The 67ms latency claim is especially relevant for real-time routing, voice systems, and high-volume automation
- –The JevBench lead is promising, but developers should validate calibration, edge cases, and benchmark methodology on their own workloads
DISCOVERED
1h ago
2026-10-08
PUBLISHED
1h ago
2026-10-08
RELEVANCE
AUTHOR
tom_w_hamer