AI News Feed
Market watch
Cramer Sticks With Monthly S&P 500 Investing Despite Market RisksImporting a Phone to the US Adds Carrier, Software and Warranty HurdlesAmazon in Talks to Buy AI Startup Decart for as Much as $7 BillionOpenAI Dot vs. Meta Muse: Price, Permissions and Availability Divide Always-On AI AgentsNew Constructs Calls Anthropic IPO 'Most Ridiculous of 2026' as $2 Trillion Valuation Draws ScrutinySouth Korea's October 1-10 exports double on chip upcycle, hitting record $26.1 billion
Cramer Sticks With Monthly S&P 500 Investing Despite Market RisksImporting a Phone to the US Adds Carrier, Software and Warranty HurdlesAmazon in Talks to Buy AI Startup Decart for as Much as $7 BillionOpenAI Dot vs. Meta Muse: Price, Permissions and Availability Divide Always-On AI AgentsNew Constructs Calls Anthropic IPO 'Most Ridiculous of 2026' as $2 Trillion Valuation Draws ScrutinySouth Korea's October 1-10 exports double on chip upcycle, hitting record $26.1 billion
Companies

Cloudflare Releases Clef-omni, a 30B-A3B Multimodal Decision Model for Text, Images, Audio and Video

Cloudflare released Clef-omni on Oct. 9, 2026, an open-weight 30B-A3B MoE model that returns probability-scored decisions across text, images, audio and video in one call.

AI News Feed · 2026-10-12 09:39

Clef-omni is a 30-billion-parameter mixture-of-experts model with 3 billion active parameters, post-trained from Qwen3-Omni-30B-A3B-Instruct. It has a 64,000-token context on Workers AI, where it is available as @cf/cloudflare/clef-omni. Self-hosting has been tested on a single H200 with about 64 GB of GPU memory in bfloat16. Cloudflare describes a decision model as one that returns typed answers with probabilities rather than generated text: a user sends a state and a schema of questions, and the model returns a probability for every allowed option. Clef supports three question types: noul, choice and score, following the Jev and SystemOne API from TypeSafe AI.

Cloudflare kept the Qwen3-Omni comprehension backbone and dropped the text-to-speech output path. The model runs a single prefill pass over the full payload, syncing video frames and audio into one sequence for joint processing. A small joint schema head then scores the options, with field vectors cross-attending across the full context to compute confidence scores. Cloudflare froze the Qwen3 backbone and trained LoRA adapters, combining label-smoothed cross-entropy with Brier score calibration. On Workers AI, supported inputs include up to four PNG, JPEG or WebP images at 64 to 1,024 tokens each; up to four WAV or MP3 audio clips of 300 seconds each, at about 780 tokens per minute; up to two MP4 or WebM video clips of 60 seconds each, sampled at two frames per second; and one to 64 questions per request. Cloudflare reports median text decisions in about 130 ms, images in about 150 ms, and a 21-second video with sound in about 1.5 seconds.

In Cloudflare's Decision Index comparison against Clef, Clef-flash and Jev, Clef-omni led on intent classification with 94.8 macro-F1 on BANKING77 and 97.7 on CLINC150+OOS, and topped Amazon ESCI at 57.8. It trailed elsewhere: 69.3 case-exact accuracy on home appliances versus 97.73 for Clef-flash, 63.3 on When2Call versus 80.97 for Jev, and 47.4 on GPQA Diamond versus Jev's 78.3. It also scored 2.6 accuracy on POP909-CL and fell below Clef on all five reported TypeSafe workflow metrics, the report said. The benchmark scores came from Cloudflare's announcement and were not independently verified.

Cloudflare also reduced Clef-flash's input price to $0.038 per million tokens, below Jev's $0.042, with 24k hosted context and 256k self-hosted. Clef-omni is priced at $0.15 per million input tokens. The company moved hosted Clef to SGLang and tuned its serving stack without releasing new weights. Median latency fell from 262 ms to 152 ms at about 800 tokens, from 616 ms to 305 ms at about 3,400 tokens, and from 2,721 ms to 1,635 ms at about 16,000 tokens. Clef support is in SGLang 0.5.22. For developers, Cloudflare said the model removes the transcription and captioning steps that multimodal agent pipelines usually need.