Models, recipes, providers, and execution

locks/recipe-support-v1.json is the support authority; locks/embedding-models-v1.json contains exact model commit, declared license, architecture, dependencies, prompt, pooling, padding, normalization, safe dimensions, context, native heads, remote-code requirement, hardware estimate, evidence URLs, limitations, and unavailable reasons.

The status matrix currently contains 13 recipes: the tiny CPU embedding fixture is supported only for offline testing, while the Qwen 9B chat pilot and all eleven production recipes are experimental. Experimental status does not authorize execution: missing license, GPU smoke, clean reload, or evaluation evidence still fails closed. BGE sparse/ColBERT/hybrid and GTE sparse are later experimental work.

Chat recipe status

Canonical recipe	Track / architecture	Status	Blocking evidence
qwen3.6-27b	chat / dense	experimental	pinned license, GPU smoke, reload, evaluation
qwen3.6-35b-a3b	chat / MoE	experimental	pinned license, GPU smoke, reload, evaluation
nemotron-cascade-2-30b-a3b	chat / hybrid MoE	experimental	pinned license, GPU smoke, reload, evaluation
nemotron-3-nano-30b-a3b	chat / Mamba-transformer MoE	experimental	pinned license, GPU smoke, reload, evaluation
olmo-3.1-32b-instruct	chat / dense instruct	experimental	pinned license, GPU smoke, reload, evaluation
olmo-3.1-32b-think	chat / dense reasoning	experimental	pinned license, GPU smoke, reload, evaluation
qwen3.5-9b-pilot	chat / pilot	experimental	not a supported production recipe

Templates, assistant-only masks, precision, quantization, target modules, and effective batch are locked in the recipe/job identity; none of these statuses assert that a GPU run has passed.

Embedding recipe matrix

Recipe / model revision	Architecture and task	Prompt, pooling, padding	Dimensions / context	License and hardware status
qwen3-embed-0.6b-lora · Qwen/Qwen3-Embedding-0.6B · 97b0c6…	Qwen3 decoder bi-encoder · retrieval	Instruct…Query…, last-token, left	32–1024 · 32,768	Apache-2.0 declared; license artifact and LoRA GPU/reload evidence missing
arctic-m-v2-full · Snowflake Arctic · 95c274…	XLM-R encoder · multilingual retrieval	query/document prefixes, CLS, right	256/768 · 8,192	Apache-2.0 declared; license artifact and pinned remote-code review missing
bge-m3-dense · BAAI/bge-m3 · 5617a9…	XLM-R · dense retrieval	no prompt, CLS, right	1024 · 8,192	Metadata declares MIT, conflicting with expected Apache-2.0; reviewed license artifact required
nomic-v2-moe-native · Nomic v2 MoE · 1066b6…	NomicBERT MoE · retrieval/classification/clustering	task prefixes, mean, right	256/768 · 512	Apache-2.0 declared; license artifact and external remote-code review missing
gte-multilingual-base-full · GTE · 9bbca1…	multilingual encoder · dense retrieval	no prompt, CLS, right	768 · 8,192	Apache-2.0 declared; license artifact and pinned remote-code review missing
cpu-tiny-embedding-fixture	deterministic test encoder	fixture q: convention, mean	fixture-only	supported only for offline CI; no quality claim

All production embedding rows use L2 normalization. Precision/quantization and hardware are recipe-specific: Qwen targets LoRA GPU execution; encoder full tunes require a measured CPU/GPU probe; MoE and remote-code paths need additional evidence. Exact commits, prompts, heads, dependencies, limitations, and evidence URLs live in locks/embedding-models-v1.json and are checked against this documentation.

OpenAI and Anthropic adapters are optional peers. Provider selection requires an explicit model and API-key environment reference. Capabilities, retry/rate policy, usage and cost, raw-envelope redaction, budget, and idempotency are recorded. Estimates are user-side planning controls—not provider hard spending caps.

Local CPU execution is available for fixtures. The Phase 20–23 finetuning runpod surface provides pinned offline contracts, read-only diagnostics, planning, dry-run output, and fake lifecycle/hardening tests. Production recipes are labeled experimental, but live Pod and volume mutation still fail closed because no authorized qualification evidence or spend was produced. Spot, generic exec/log REST, direct Secrets representation, Serverless, and fleet remain unavailable. See RunPod execution and support status.