Support matrix v7

This page lists specific model variants that EDB has tested and confirmed to work with AIDB, split into local models (run on your Postgres host) and remote models (NVIDIA NIM microservices, called over the network). Each table follows the same shape: a model identifier, whether it's a default, and a description of what the model is good for — use the description to pick the right variant for your use case.

Local models

Default models

AIDB registers these models automatically when the extension is created — no configuration needed to start using them. See Default models for how to use them, and Model providers for what each provider means.

Model nameProviderDescription
bertbert_local (Candle)General-purpose English text embeddings. Good default for most text search and RAG use cases.
clipclip_local (Candle)Joint text/image embeddings for multimodal search.
t5t5_local (Candle)Lightweight text-to-text generation — translation, summarization, question answering.
llamallama_instruct_local (Candle)Instruction-following local chat/completion model.
bge-small-en-v1.5-f16llamacpp_embeddings (llama.cpp)Compact, fast English text embeddings.
nomic-embed-text-v1.5-Q8_0llamacpp_embeddings (llama.cpp)General-purpose English embeddings with a larger context window (2048 tokens) than bge-small-en-v1.5-f16.
bge-m3-f16llamacpp_embeddings (llama.cpp)Multilingual embeddings with an 8192-token context window — good for long or non-English documents.
qwen3-embedding-0.6b-Q8_0llamacpp_embeddings (llama.cpp)Larger embedding model with a 16384-token context window.
qwen3-embedding-4b-Q8_0llamacpp_embeddings (llama.cpp)Largest pre-registered embedding model (20480-token context) — higher quality, more compute.
qwen3.5-0.8b-Q8_0llamacpp_generate (llama.cpp)Small, fast chat/completion model with a very large context window (262144 tokens).
llama-3.2-1b-instruct-Q8_0llamacpp_generate (llama.cpp)Compact Meta Llama 3.2 instruction model for lightweight local chat/completion.
dummydummyReturns fake data — for testing pipelines without running real inference.

Other tested variants (Candle)

The following HuggingFace model variants have been tested and verified to work with AIDB's Candle-based local providers, beyond the defaults above. Use any of these identifiers with the corresponding config helper when calling aidb.create_model() — see Custom models.

BERT (bert_local)

Model identifierDefaultDescription
sentence-transformers/all-MiniLM-L6-v2General-purpose English sentence embeddings; fast and small. This is what the default bert model uses.
sentence-transformers/paraphrase-multilingual-mpnet-base-v2Multilingual embeddings — use when your text isn't (only) English.
sentence-transformers/multi-qa-MiniLM-L6-cos-v1Tuned for question-answering and retrieval, rather than general sentence similarity.
sentence-transformers/paraphrase-TinyBERT-L6-v2Smaller and faster than the default, at some quality cost.
SELECT aidb.create_model('my_bert', 'bert_local',
    config => aidb.bert_config('sentence-transformers/paraphrase-multilingual-mpnet-base-v2'));

All BERT variants support text embedding (aidb.encode_text()).

CLIP (clip_local)

Model identifierDefaultDescription
openai/clip-vit-base-patch32The only tested CLIP variant. Joint text/image embeddings. This is what the default clip model uses.

CLIP supports both text embedding (aidb.encode_text()) and image embedding (aidb.encode_image()).

T5 (t5_local)

Model identifierDefaultDescription
t5-smallFastest, smallest T5 variant. This is what the default t5 model uses.
t5-baseLarger than the default — better quality, slower.
t5-largeLargest T5 variant — highest quality, most resource use.

T5 variants support both text encoding (aidb.encode_text()) and text generation (aidb.decode_text()).

Llama (llama_instruct_local)

Model identifierDefaultDescription
TinyLlama/TinyLlama-1.1B-Chat-v1.0Small, fast instruction-following chat model. This is what the default llama model uses.
HuggingFaceTB/SmolLM2-135M-InstructExtremely small and fast — lower quality, lowest resource use of these variants.
HuggingFaceTB/SmolLM2-360M-InstructSmall — a step up in quality from the 135M variant.
HuggingFaceTB/SmolLM2-1.7B-InstructLargest of these variants — better quality than the default, at higher resource use.

All Llama variants support text generation (aidb.decode_text()).

Note

These are the pre-vetted variants confirmed to work with AIDB. You aren't limited to these — any compatible HuggingFace model can be used with the appropriate config helper, but only the variants listed here have been tested by EDB.

Remote models (NVIDIA NIM)

AIDB supports NVIDIA NIM microservices via several providers — see External model connections for how to register one. All other models can be self-hosted using KServe and registered with AIDB using aidb.create_model().

For the full list of supported NIM models per provider (completions, embeddings, CLIP, OCR, and reranking), including which are pre-deployed by EDB Hybrid Manager's Model Serving, see GPU Recommendations for Default NIM Models.