Featherless releases Simple Jev, turning open models into zero-shot classifiers

Serverless inference provider Featherless has released Simple Jev, an open-source library that converts open-source AI models into high-speed, zero-shot classification engines.

Simple Jev builds on this month's launch of Jev, the closed-source, text-only system from TypeSafe, and extends structured decision functionality to open-source models. Applications can return categorical assignments or binary choices without generating conversational text, and on Featherless's hosted endpoints the library handles images too.

Don't use a tank to deliver a pizza

Eugene Cheah, Featherless's CEO and co-founder, told The New Stack that reaching for a bleeding-edge frontier model to classify something as ordinary as a support ticket amounts to "using a tank to deliver a pizza" in terms of tooling overload.

"Sure, the pizza tank gets there eventually, but it's slow, it's expensive, and it's the wrong vehicle," Cheah said, adding that businesses need fast reflexes and that the industry neglected that kind of AI for too long.

He concedes that large closed-source models can handle such tasks, but argues that they answer by generating text through "enormous multi-trillion-parameter LLMs" at frontier prices, making them the wrong tool. Simple Jev never writes an answer: it stops the model where it would choose, reads the score of each allowed option and outputs the probabilities.

Vision support for decision workflows currently works only with Gemma or Qwen models.

What a zero-shot classification engine does

Zero-shot engines use pre-trained language models to sort inputs into categories they were never explicitly trained on, relying on transfer learning instead of task-specific data. Cheah notes that classifiers are not new, and that universities taught the approach before ChatGPT arrived; what is new is a well-designed zero-shot API.

He says any graduating AI/ML PhD could reimplement Simple Jev from a single sentence by building a "shared-prefix, two-stage, prefill-only, logit-based classifier", and expects hundreds of open-source Jev clones now that the code is public on GitHub.

Pricing and the open-model argument

Featherless is also opening free public endpoints that need no API key or login; the demo is capped at 2,000 tokens of context and two requests per second. Production pricing starts at $0.03 per million input tokens with output tokens free, against $0.042 for Jev.

Cheah estimates a typical decision over an image or a paragraph at 500 to 1,200 tokens, which at $0.03 per million is roughly three thousandths of a cent, or about $15 to $35 per million decisions.

Featherless urges teams to move on from "monolithic generalist models" and predicts millions of lightweight, dedicated models will drive the future; it cites OpenAI's CLIP from 2021 and Microsoft's Florence-2 from 2024.