Granite Switch 4.1 vs Jev

Granite Switch 4.1

One self-hosted Apache 2.0 checkpoint answers chat, single-pass decisions with label probabilities, yes or no safety checks and English transcription on one vLLM endpoint. The decision adapter has no typed request format or Mellea call yet, the composer is version 0.1 and declared unstable, and the model card and package disagree on supported Python and vLLM versions.

Jev

Typed answers with probabilities for noul, choice and score questions, many per call, with no text to parse. Early access behind a waitlist, with no free tier or free credits found.

Granite Switch 4.1

1. Pass chat_template_kwargs: {"adapter_name": "decision"} or "guardian-core". Without it the request is ordinary chat on the base model
2. For a decision, send the state and the ### DECISION question as two user turns, request logprob_token_ids for the option labels, and set max_completion_tokens to 1
3. Use single-token option labels. Split more than 128 labels into chunks with the identical prompt and apply one softmax over all returned logprobs
4. Parse guardian-core answers as JSON, {"score": "yes"} or {"score": "no"}, where yes means the criterion is met
5. Add authentication before exposing the vLLM server. Neither vLLM nor the checkpoint requires a key by default

Jev

1. Put every independent question about one state into a single call. They run in parallel and the state is billed once
2. Pin jev-1.13.0 instead of jev-latest once you've tuned confidence thresholds
3. Back off exponentially on 429 and 529. The limits move with demand
4. Keep state to what the decision needs. Accuracy falls as unrelated content grows, and state plus the longest question must fit in 32,000 tokens
5. Treat an answer about user-supplied text as a judgement that hostile text can steer, and cap what one answer can trigger