# Cloudflare Launches Clef-Omni for Native Multimodal AI Decision Routing Open-weight decision family adds audio and video alongside major speed and pricing upgrades. - **Source:** [Cloudflare Blog](https://blog.cloudflare.com) - **Published:** 2026-10-09 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/09a1b39e-bf2a-4b02-86e1-8a0debe1a2b5 ## Unifying Multimodal Decision Models Cloudflare expanded its open-weight decision model family by launching Clef-omni. Unlike standard decision models that rely solely on text or isolated image frames, this release natively ingests text, images, audio, and video payloads simultaneously. It eliminates complex cascading pipelines previously required for multi-source automated workflows. ## Qwen3 MoE Backbone Architecture Clef-omni is built upon a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation. Cloudflare retained the core comprehension backbone while stripping out text-to-speech generation components. The model runs a fast prefill pass across the entire payload, synchronizing video frames with audio streams directly in a single unified sequence. ## Two-Stage Schema Constrained Scoring Instead of generating autoregressive output tokens like standard LLMs, Clef-omni extracts candidate values from internal embeddings. A two-stage attention routing gathers evidence across modalities before field vectors cross-attend for confidence scoring. Built-in lexical grammar enforces schema constraints while maintaining high semantic accuracy across various prompt structures. ## Fast Low-Latency Inference Metrics Skipping transcription and captioning enables significant latency improvements across diverse payloads. Median text decisions process in 130 milliseconds, image inputs take roughly 150 milliseconds, and audio files finish within a few hundred milliseconds. A full 21-second video with synchronized sound completes scoring in approximately 1.5 seconds. ## Competitive Benchmarking Results Clef-omni delivered competitive accuracy against alternatives like Jev on standard evaluation benchmarks. It scored 98.2 on BFCL case exact, 92.7 on API-Bank accuracy, and 94.8 macro-F1 on BANKING77. For domain workflows, it achieved 82.0 on invoice processing primary action and 71.6 on customer service tasks. ## Clef-Flash Price Reductions Cloudflare reduced the hosted price of Clef-flash from $0.09 down to $0.038 per million input tokens. To achieve this cost structure, Cloudflare reduced the hosted context window from 64k to 24k tokens after telemetry revealed only 0.24% of production requests exceeded that specific threshold. ## Infrastructure Speedups Via SGLang The standard Clef model received substantial throughput improvements through serving infrastructure optimizations on Workers AI rather than architectural modifications. Integrating SGLang in pull request 42721 enabled up to a 2.0x median speedup on 3,400-token payloads, reducing latency from 616 milliseconds down to 305 milliseconds. ## Production Deployments Across Cloudflare Internal engineering teams already deploy Clef models across multiple production systems without custom fine-tuning. Production use cases include automated spam issue closure on public GitHub repositories, phishing moderation within the EmDash CMS plugin library, PII detection for data loss prevention, and malicious domain detection in threat intelligence. ## Open Weights and API Integration The complete open weights for Clef-omni, Clef, and Clef-flash are available on Hugging Face with full 256k self-hosting support. Hosted endpoints maintain drop-in compatibility with the Jev API format. Developers can invoke multimodal decision evaluation directly through Workers AI Gateway by modifying the target model identifier. ## Sources covering this story - https://blog.cloudflare.com/clef-faster-cheaper-multimodal/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._