Clef: Open-weight decision models, and new RL fine-tuning platform
Cloudflare released Clef and Clef-flash, open-weights decision models hosted on Workers AI alongside a new reinforcement learning fine-tuning platform.
summary is enough The post details Cloudflare's new decision models and RL platform, but the core technical approach and early user benchmarking are what matter most.
summary
Cloudflare introduced Clef and Clef-flash, open-weights decision models built on Qwen backbones designed for fast, probabilistic structured classifications in agentic workflows. Instead of generating text autoregressively, the models use a non-autoregressive two-stage attention routing process to score schema choices directly. Alongside the models, Cloudflare debuted an RL fine-tuning platform leveraging its AI Gateway, containers, and Workers AI infrastructure. Commenters debated pricing, the open-weights distinction, and whether decision models are simply a clever repackaging of existing classification techniques.
key points
- Clef models utilize Qwen backbones, a 64k context window, a vision encoder, and specialized non-autoregressive scoring heads for fast structured outputs.
- Input pricing is $0.24 per million tokens for Clef and $0.09 for Clef-flash, which is higher than some competitors.
- The accompanying reinforcement learning platform uses Cloudflare AI Gateway, Workers AI, and container sandboxes for custom fine-tuning.
who should read it
AI engineers and developers building automated agentic workflows that require rapid, structured classification decisions.
discussion positive
consensus
Commenters generally view decision models as a useful, accessible optimization rather than a groundbreaking new paradigm, noting they function as specialized classification heads on base LLMs.
pushback
Several users criticized Clef's pricing as higher than competitors like Jev and reported mixed results in early practical benchmarks.
It's not a "new paradigm", it's a low-hanging fruit that's been lying around for years; Typesafe were the first to bother to stop and pick it up, and market the shit out of it. But it was still a low-hanging fruit.
There are many, many of those left around, because AI frontier is moving forward so fast, everyone is racing ahead. Which is why I laugh when people say AI is not transformative and LLMs are a dead end (and my favorite, "what are we going to do with all those GPUs when the bubble pops?"). Even if SOTA LLMs hit a hard capability limit tomorrow and never advanced again, there's a …
It contextualizes the sudden flood of decision models as low-hanging fruit left along the fast-moving AI frontier.
Generated by an LLM from the article and thread. Check the sources before relying on it.