AOI v4 Piggyback LLM Reference Library — 24-Model Survey (Closed API + Open-Weight), Confirming the LLM-Free Confidence Gate

Reference survey only. AOI v4's confidence gate resolves abstentions without calling any of these models; no API/token spend occurs in the current pipeline. See related identifier for the AOI v4 engine itself (10.5281/zenodo.21368154). Reference docs for 24 LLMs, organized as an optional "piggyback" resource for AOI v4 — none of these are runtime dependencies of AOI v4, which remains numpy-only and needs no LLM to run. This library exists so AOI v4's confidence-gate abstentions have somewhere to optionally escalate to later, and so open-weight models can be pulled locally on demand. Compiled from public provider documentation (pricing pages, model cards, official docs) — this is a documentation survey, not original research or a novelty claim. Closed / API-only (no local weights possible): Claude Fable 5, Claude Sonnet 5, Claude Opus 4.8 (Anthropic); Gemini (Google); Grok (xAI). Open-weight (downloadable from Hugging Face): Qwen (Alibaba), Llama (Meta), Mistral (Mistral AI), DeepSeek, Gemma (Google), Phi (Microsoft), Falcon (TII), Yi (01.AI), Command-R (Cohere), StableLM (Stability AI, inactive since Apr 2024), MPT (MosaicML/Databricks, superseded by DBRX), BLOOM (BigScience), Pythia (EleutherAI), OLMo (Allen Institute for AI), Jamba (AI21 Labs), GLM (Zhipu AI/Z.ai), Baichuan, InternLM (Shanghai AI Lab), Aya (Cohere For AI). Each card documents provider, license, context window, capabilities, and (where applicable) API pricing, plus a note on how that model could optionally plug into AOI v4's confidence gate as an escalation layer for abstained cases — strictly hypothetical, never a dependency. Licensing varies widely across the open-weight set: Apache 2.0/MIT (Qwen, DeepSeek, Phi, Yi, GLM, InternLM, OLMo — cleanest for commercial use) vs. non-commercial or usage-gated (Command-R, Aya Expanse/Vision, Jamba 1.5+, Falcon).