Local AI model portfolio¶
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.
| Role | Default candidate | Envelope | Promotion suite |
|---|---|---|---|
text_fast |
Qwen3 0.6B Q8 | 1 GB / 4K | text_instant |
text_quality |
Qwen3 1.7B Q5 | 2 GB / 8K | text_ingestion |
text_scholar |
Qwen3 4B Q4 | 4 GB / 16K | text_enrichment |
text_cjk |
Qwen3 4B Q4 | 4 GB / 8K | text_multilingual |
embedding_search |
EmbeddingGemma 300M | 1.5 GB / 2K | embedding_retrieval |
function_routing |
FunctionGemma 270M | 1 GB / 4K | function_routing |
multimodal_analysis |
Gemma 4 E2B | 12 GB / 32K | multimodal_analysis |
audio_fast |
Whisper small | 768 MB | audio_fast |
audio_english |
Distil-Whisper large-v3 | 2 GB | audio_english |
audio_multilingual |
Whisper large-v3-turbo | 2 GB | audio_multilingual |
audio_translation |
Whisper medium | 2 GB | audio_translation |
The Qwen ladder is small-first. EmbeddingGemma is a separate vector capability. FunctionGemma is experimental and not a general dialogue model. Gemma 4 E2B is an experimental Safetensors text/image/audio candidate, not an LLamaSharp GGUF model. Turbo performs source-language transcription; Whisper medium remains the speech-to-English translation baseline.
Sources: Qwen3 GGUF, EmbeddingGemma, FunctionGemma, Gemma 4 E2B, Whisper, and Distil-Whisper.
model_catalog owns provenance, license, checksum, capabilities, compatibility, and gates. operational_roles owns workload envelopes. role_requirements owns objective promotion policy. The enum-backed models section remains the executable bridge for currently integrated LLamaSharp/Whisper roles. Automatically downloadable executable artifacts are SHA-256 pinned; gated Google artifacts have no automatic download until an administrator accepts their terms and installs a verified artifact.