Simple-jev by featherless AI – Turn context into clear decisions

Featherless AI released Simple-jev, an open-source tool that turns existing open language models into classifiers by reading next-token logits for allowed labels, returning structured JSON with choices, confidence scores, and probabilities. The demo API runs the featherless-ai/gemma-4-26B-A4B-classifier model with limits of 2,000 tokens and 4 requests per second, and image context is supported only for Gemma and Qwen models. Simple Jev models on Featherless start at $0.03 per million input tokens, with the service currently in beta and hosted fine-tuned models and fine-tuning planned as usage grows.

Bring the whole conversation. We support chat-format context. Send a conversation as messages with roles, then ask questions about the exchange—not just the last message. Explore chat context docs.html context OPEN MODELS. STRUCTURED DECISIONS. Turn context into clear decisions. Give any open model a message and a few questions. Get choices, scores, and answers your code can use. Source on Github here https://github.com/featherless-ai/simple-jev A little model. A very specific job. curl https://simple-jev-demo-api.featherless.ai/v1/classifier \ -H 'Content-Type: application/json' \ -d '{ "model": "featherless-ai/gemma-4-26B-A4B-classifier", "state": "Mia owns a red bicycle.", "questions": { "color": { "type": "choice", "instructions": "What color is Mia’s bicycle?", "criteria": {"red": null, "blue": null} } } }' For production usage and higher rate limits, please signup for a developer account at https://featherless.ai { "color": { "type": "choice", "choice": "red", "confidence": 1, "probabilities": { "red": 1, "blue": 1.8874485308018052e-10 } } } Actual Gemma response, showing the answers field. Results may vary. Demo limits: 2k tokens · 4 requests/second. Full API documentation ↗ docs.html 01 / TRY IT Edit scenarios, build questions, and see real model responses in the interactive playground. SEE IT IN ACTION MORE WAYS TO BRING CONTEXT Use images as context for your questions. Image and vision support is available for Gemma and Qwen models only . 02 / THE IDEA Use existing open language models as classifiers. No separate classifier head required. A message, chat history, or structured state. Ask several questions about the same input. We read the next-token logits for the allowed labels and reuse the shared prompt cache across questions. Your application receives structured JSON. Route a ticket, rank a result, or choose the next step. ONE MORE THING Really Fancy Decision Training. Bring examples and answers—or ask a larger teacher model to label them. RFDT fine-tunes the decisions your application actually needs. READY FOR MORE? Simple Jev models on Featherless start at $0.03 per million input tokens , with higher limits on developer plans. Currently in beta. Support for hosted fine-tuned models and fine-tuning is on the way as usage grows. Prices may change after beta. Refer to Featherless.ai official pricing https://featherless.ai/ for up-to-date information.