Published
SimpleJev turns an open model into a classifier by reading the next-token logits
Featherless posted SimpleJev, an Apache 2.0 server that scores Choice, Score, and Noul from a Hugging Face model's next-token logits. A public demo needs no key. JevBench listed Qwen3.8-27B at 67.3, with hard-tier accuracy 75.0%, matching Jev on that slice and trailing on the composite.
Featherless posted Simple Jev on September 18. The repo is featherless-ai/simple-jev, Apache 2.0. The site is simple-jev.featherless.ai.
You send shared state, or a chat history, plus named questions. The server prefills the common prefix once, reads the next-token logits for each allowed label, and builds the JSON without sampling a completion. POST /v1/classifier is the path. /v1/systemone is an alias of the same handler. Choice, Score, and Noul are the three question types. The README says the project does not reproduce TypeSafe’s model architecture or training.
A public demo at simple-jev-demo-api.featherless.ai needs no key. The README and JevBench both cap it at 2,000 tokens of context and 2 requests per second. The marketing page also printed 4 RPS. The sample model id is featherless-ai/gemma-4-26B-A4B-classifier. Hosted Simple Jev on Featherless is listed at $0.03 per million input tokens, in beta.
Florian S added the rows to JevBench on September 21. His v1.2.6 post put SimpleJev Qwen3.8-27B at rank 9. The table we read is v1.2.8, after more systems landed: that row is 67.3, rank 13. Intelligence 89.7, close to Jev’s 90.4. Hard-tier accuracy 75.0%, next to Jev 74.1%. Cost 39.5, about $0.104 estimated per 1,000 decisions, is the weak axis. Speed uses the public-demo x2 load adjustment (1.01 s raw, 2.03 s adjusted). SimpleJev Qwen3.6-35B-A3B is 63.8, rank 24.
The Hugging Face server in the repo is text only. The site ships separate Gemma and Qwen vision demos. RFDT, in the same repository, fine-tunes a student on answer-token logits from your labels or a teacher. We did not call the demo or train a student.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Featherless posted Simple Jev on September 18. The README (featherless-ai/simple-jev, Apache 2.0, Copyright 2026 Featherless AI / Recursal AI) is the spec. POST /v1/classifier scores shared state or chat messages; /v1/systemone is an alias of the same handler. The server reads next-token logits for the allowed labels and builds the JSON. It does not generate a completion. Public demo at simple-jev-demo-api.featherless.ai: no login, 2k-token context, 2 requests per second in the README and in JevBench's note (the marketing page also printed 4 RPS). Florian's September 21 post put SimpleJev Qwen3.8-27B at rank 9 on v1.2.6. The table we read is v1.2.8: that row is 67.3, rank 13, Intelligence 89.7, Calibration 81.1, Speed 71.2, Cost 39.5, hard-tier 75.0% against Jev 74.1%, p50 1.01 s raw then 2.03 s with the public-demo x2 adjustment, about $0.104 estimated per 1,000 decisions. SimpleJev Qwen3.6-35B-A3B is 63.8, rank 24, hard-tier 66.4%. Cost is a hosted size-class estimate, not a Featherless bill. The README says this does not reproduce TypeSafe's architecture or training. The HF server is text only; the site's vision demos are Gemma and Qwen on a separate path. Hosted Simple Jev on Featherless is listed at $0.03 per million input in beta. We did not call the demo. TypeSafe's Master Customer Agreement section 2.3(f) forbids publishing benchmarks of the Services; the JevBench rows are reported as published.
Compare
SemIf also reads option logits, from a frozen Qwen3.5-4B, and leads the open composite at 74.7. SimpleJev is a server you point at any compatible Hugging Face chat model. On JevBench the Qwen3.8-27B demo matches Jev's hard-tier accuracy (75.0% versus 74.1%) and loses the score on cost and speed. djev is a DiffusionGemma API with native images and a 74.3 composite. Kev and jeff copy POST /v1/systemone with a trained head. SimpleJev's /v1/systemone is an alias of /v1/classifier. RFDT in the same repo fine-tunes answer-token logits on your labels; we did not train it.
Terms
- SimpleJev
- Featherless's Apache 2.0 classifier server. It reads next-token logits for allowed labels and returns Choice, Score, or Noul JSON. JevBench v1.2.8 lists Qwen3.8-27B at 67.3.
- Next-token logit read
- Score the allowed answer tokens from one prefill, with no decode loop. SimpleJev reuses the shared prompt's KV cache across questions in the same request.