Model and provider choice

Require task family, input/output contract, quality threshold, latency/SLA, hardware, budget, privacy boundary, context length, license constraints, and whether the use is inference or training.

Workflow

1. Enumerate candidate model IDs with exact revision, tokenizer, architecture, context limits, modality, license, and availability.
2. Check task-family fit and training compatibility: tokenizer, adapter target modules, quantization, sequence length, image processor, and framework support.
3. Compare local OpenAI-compatible prefixes, HF Router/provider routes, and direct local inference without implying that a prefix proves model identity.
4. Record model-ID freshness, provider health, cost and latency assumptions, quality evidence, and fallback behavior.
5. Produce a reproducible model decision record with rejected alternatives and hand it to ml-stack-training, ml-stack-evaluation, or ml-stack-deployment.

Boundary

Do not download, publish, route paid traffic, or mutate provider settings without approval. Do not infer license permissions from a model card summary. Abstain when revision, license, tokenizer, or quality evidence is missing. Next owner: ml-stack-training for fine-tuning, ml-stack-evaluation for benchmark confirmation, and ml-stack-deployment for serving.

See model selection, local models, gateway and cost, and revision and license.