Granite Switch 4.1 vs ScaleDown

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.

ScaleDown

Classify answers Jev-shaped choice, noul and score questions with probabilities, and Extract returns values with confidences and character offsets, at $0.05 per million input tokens with output free. There is no status page, no published rate limit and no public changelog, and the site's compliance claims contradict each other.

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

ScaleDown

1. Send several questions about one item in one /v1/scaledown call. Each becomes its own Classify call and the usage is summed
2. Pin classify-1, extract-1 or summarize-1 once you've tuned thresholds on your own labelled cases
3. Set context_chars to 0 and use top_n on Extract unless you need the 500 characters of context each side
4. Set reasoning: false on /classify when you only need the label. The reasoning call bills its input at 2x
5. Back off and retry on 429, 502 and 504. No limits are published and calls have no side effects