Cloudflare open-sources Clef decision models and debuts an RL fine-tuning platform

Seriously, how did so many people ship decision models within days of Jev dropping? Was this brewing for a while, or is it just trivially easy to copy?

The concept was around a year before Jev, go look up Laya. None of this is new.

You don't need the high-scale compute, clean data, and RL pipeline an LLM needs. Take an already trained transformer and fine-tune it into one.

Small models could do this for a while, just slower. I've had a Qwen 0.8b doing basic image classification in under 500ms on my laptop for ages. The real insight was watching the reaction to Jev and seeing there's market interest in selling it as its own thing.

Transformers already output a probability distribution. For ChatGPT that's next-token prediction, but it can just as easily be a list of classes. Jev mostly innovated on the API and the product concept, and copying an API is easy when any pretrained LLM can be adapted.

Jev nailed the ergonomics, an intuitive API with structured data. They'd say their edge is model intelligence and calibration, which is the genuinely hard part. There's no free lunch with these things.

If your use case is narrow you can train a BERT-based classifier on a laptop in an hour. It answers faster than a roundtrip to Clef or Jev and uses under 1GB of memory.

Getting training data that works for calibrated classification is hard, and I hear conflicting reports on how calibrated any of these actually are. Calibration barely matters though when you're replacing people who were reading softmax probabilities off an LM head. Accuracy is what they care about.

The real question for AI in automation is whether it decides consistently and predictably, near 100% determinism. Plenty of people hit that same wall and started working out the answer.

Cloudflare is earning real goodwill from me. Consistently interesting releases at great prices. Almost too good to be true.

Good to see Cloudflare shipping edge-level models for normal consumers.

Pricing is $0.24 per million input tokens, roughly 6x Jev. Clef-flash at $0.09 is way more competitive. Odd that their pareto frontier chart leaves cost out entirely.

brb rebuilding my entire cloud stack on Cloudflare.

It takes image input. That's the part I care about.

That's probably driven by their own need to feed the model screenshots of emails and webpages to catch phishing even when the HTML is obfuscated.

I'm not an ML person at all, is a decision model really easy enough to build that there's already a flood of Jev alternatives?

You can make a basic one in minutes off an existing open-source model. Latency won't be great but it works, you just force the structured output to a schema. It's not new technology, it's a new use case.

Yeah, it's very easy if you have basic ML knowledge.

The basics are simple, and depending on your need the model can be tiny and run on a phone, thousands of requests in under 100ms. I've been playing with this for a year: a personal email classifier, a couple of classifiers that play Doom, and now a request router that picks a classifier and falls back to an LLM. Jev shaped the concept well, and now everyone is rushing the hype window.