Clef: Open-source decision models, and new RL fine-tuning platform
6 hours ago
- Cloudflare releases Clef and Clef-flash, open-source decision models hosted on Workers AI, with Apache 2.0 licenses on Hugging Face.
- Clef leads the Jev Decision Index benchmark, offering faster, smarter, and fully Jev-API-compatible decision-making models.
- Decision models return typed structured outputs with probabilities for routing, escalation, or human deferral, enabling autonomous agent workflows.
- Clef features a vision encoder, 64k context window, and lower latency compared to Jev and general LLMs (e.g., 2.2s vs 4.7s for domain classification).
- Clef is built on Qwen with a non-autoregressive architecture using two-stage attention routing and schema-bound scoring, trained with label-smoothed cross-entropy, Brier loss, and RLCD.
- Cloudflare introduces a reinforcement learning (RL) fine-tuning product and service, using AI Gateway, Workers AI, Containers, and Trainer to customize Clef for specific workloads.
- Fine-tuning enables specialized use cases like Trust & Safety, support triage, and bot classification, trading general performance for domain accuracy.
- The models are enterprise-ready with no data retention or training on customer requests, and available via developer docs and live demo benchmarks.