Cloudflare has introduced Clef-omni, a significant new addition to its open-weight Clef family of decision models. Launched on October 9, 2026, this advanced model analyzes and scores text, images, audio, and video inputs in a single, unified API request. This integrated processing capability streamlines multimodal AI agent development by eliminating separate transcription or captioning steps. Concurrently, Cloudflare announced strategic enhancements, including a price reduction for Clef-flash and a notable speed increase for its hosted Clef service.
Understanding Multimodal Decision Models
Clef-omni operates as a decision model, providing typed answers with associated probabilities rather than producing free-form text. Users present a state alongside a schema of questions, and the model then assigns probabilities to each permissible option. It supports three core question types—noul (true or false), choice, and score—aligning with the API format established by TypeSafe AI’s Jev and SystemOne models.
Technical Profile and Performance
Clef-omni features 30 billion total parameters with 3 billion active parameters configured as a Mixture-of-Experts (MoE). It supports an extensive 64,000-token context window on Cloudflare’s Workers AI platform. For self-hosting, validation occurred on a single H200 GPU, requiring approximately 64 GB of GPU memory. While excelling in intent classification, Clef-omni shows comparatively lower accuracy on complex reasoning and workflow benchmarks. It achieved a 97.7 macro-F1 score on CLINC150+OOS, but recorded 47.4 on GPQA Diamond, contrasting with Jev's 78.3.
Operational Mechanism and Input Capabilities
Developed through post-training from the Qwen3-Omni-30B-A3B-Instruct foundation, Clef-omni retains Qwen3's robust comprehension backbone while removing its text-to-speech output pathway. The model processes the entire multimodal payload in a single prefill pass, synchronizing video frames and audio data into one continuous sequence for joint analysis. Cloudflare utilized a training methodology combining label-smoothed cross-entropy with Brier score calibration, employing LoRA adapters on a frozen Qwen3 backbone. On Workers AI, it supports up to four images (PNG, JPEG, WebP, 64-1,024 tokens each), four audio clips (WAV, MP3, 300 seconds each), and two video clips (MP4, WebM, 60 seconds each). Processing speeds are impressive: text decisions in about 130 milliseconds, images in roughly 150 milliseconds, and a 21-second video with sound in approximately 1.5 seconds.
Benchmark Overview and Ecosystem Enhancements
Cloudflare’s internal comparisons highlight Clef-omni’s strength in intent classification, leading with a 94.8 macro-F1 on BANKING77 and 97.7 on CLINC150+OOS. However, its performance on other benchmarks indicated areas for growth, such as 69.3 accuracy for home appliances versus Clef-flash's 97.73, and 63.3 for When2Call trailing Jev’s 80.97. Beyond Clef-omni, Cloudflare optimized existing services. Clef-flash pricing was reduced to $0.038 per million tokens, making it more competitive than Jev's $0.042. Hosted Clef also saw a dramatic 1.7x to 2.0x speed increase, with median latency for 800 tokens dropping from 262 ms to 152 ms, thanks to SGLang migration and serving stack tuning, integrated into SGLang 0.5.22.
Key Advantages
- Clef-omni provides unified decision-making across text, images, audio, and video in a single call.
- This open-weight, 30B-A3B MoE model is based on Qwen3-Omni, licensed under Apache-2.0, with competitive pricing.
- Cloudflare enhanced its AI ecosystem through Clef-flash pricing reductions and substantial speed improvements for hosted Clef.
- The model excels in intent classification tasks but trails in more complex reasoning and workflow evaluations.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
