LibraryLocal AI model portfolio
In this page
Understand the local AI model catalog, role boundaries, runtime compatibility and promotion gates.
Where this lives in the code
config/ai.jsonsrc/MediaEngine.AI/Configuration/AiResourceProfileCatalog.cssrc/MediaEngine.AI/Configuration/AiModelCatalogDefaults.cssrc/MediaEngine.AI/Configuration/AiSettings.cssrc/MediaEngine.AI
Tuvima Library separates a model artifact, its operational role, and the product feature using that role. This prevents an embedding model from becoming a chat model and keeps experimental runtimes out of the production GGUF lifecycle.
| Resource profile / pack | Supported artifact | Declared download | Memory envelope | Role use |
|---|---|---|---|---|
| Essential | Qwen3 0.6B Q8_0 | 639 MB | 1,024 MB | All four logical text roles |
| Standard | Qwen3 1.7B Q5_K_M | 1,260 MB | 2,048 MB | All four logical text roles |
| Advanced | Qwen3 4B Q4_K_M | 2,500 MB | 4,096 MB | All four logical text roles |
| Optional audio pack | Whisper Medium | 1,500 MB | 2,048 MB | audio only |
AiResourceProfileCatalog owns the selected artifact and clones its definition into text_fast, text_quality, text_scholar, and text_cjk. The roles have different context/output budgets while sharing the same file, checksum, and download. The committed profile is Standard; hardware eligibility can lower the effective profile. The audio pack is disabled in the committed configuration.
AiModelCatalogDefaults contains only these supported launch entries. EmbeddingGemma, FunctionGemma, Gemma multimodal, alternative Whisper models, and separate embedding/function/multimodal roles from earlier portfolio designs are not current selectable catalog entries. Such experiments require explicit implementation and validation before promotion.
Sources of runtime truth are src/MediaEngine.AI/Configuration/AiResourceProfileCatalog.cs, AiModelCatalogDefaults.cs, and AiSettings.cs. Model provenance URLs, checksums, capabilities, and validation objectives are recorded there.
The code-owned catalog records provenance, license, checksum, capabilities, compatibility, and gates. Code-owned role requirements record validation objectives. Models, ModelCatalog, and RoleRequirements are JSON-ignored runtime properties; config/ai.json selects the resource profile and optional pack rather than redefining them. Automatically downloadable executable artifacts are SHA-256 pinned.