Cloudflare's Clef family of decision models just got three upgrades in one release: a multimodal Clef-omni, a cheaper Clef-flash, and a faster Clef.
Clef-omni adds audio and video input to the existing Clef lineup, which already handled text and images, running on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation with the text-to-speech output layer stripped out since Clef scores options rather than generating text. Because it skips token generation, Clef-omni returns text-only decisions in about 130 milliseconds, image-based ones in about 150 milliseconds, and a full 21-second video clip with audio in about 1.5 seconds, all through a single API call priced at $0.15 per million input tokens. Clef-flash's price dropped from $0.09 to $0.038 per million input tokens, undercutting TypeSafe's Jev, though its hosted context window shrank from 64k to 24k tokens to make that price work. Clef itself got faster too, up to 2x at the median, through a switch to SGLang serving infrastructure, with no changes to the underlying model weights.
The pitch here is narrower than a chatbot upgrade: Cloudflare bills these as decision models, not LLMs, classifiers that pick from a fixed set of options instead of generating open-ended text, which is why they can digest audio and video without a transcription step in front of them. That's a real efficiency argument for agentic workflows like routing, tagging, or intent detection, where schema-constrained speed matters more than fluent prose.
Cloudflare's own benchmarks show Clef-omni losing ground to the non-multimodal Clef on several classification tasks, including a notable drop on home appliance categorization - a reminder that for a model line built, trained, and launched within about a week, adding modalities isn't free.