Support matrix ============== This page lists specific model variants that EDB has tested and confirmed to work with AIDB, split into :ref:`Local models ` (run on your Postgres host) and :ref:`Remote models ` (called over the network — NVIDIA NIM microservices, plus OpenAI and Anthropic model families for AIDB’s native agentic adapters). 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 :ref:`Default models ` for how to use them, and :ref:`Model providers ` for what each provider means. .. csv-table:: :header: Model name,Provider,Description :widths: 12,15,25 :align: left :class: longtable `bert`,`bert_local` (Candle),General-purpose English text embeddings. Good default for most text search and RAG use cases. `clip`,`clip_local` (Candle),Joint text/image embeddings for multimodal search. `t5`,`t5_local` (Candle),"Lightweight text-to-text generation — translation, summarization, question answering." `llama`,`llama_instruct_local` (Candle),Instruction-following local chat/completion model. `bge-small-en-v1.5-f16`,`llamacpp_embeddings` (llama.cpp),"Compact, fast English text embeddings." `nomic-embed-text-v1.5-Q8_0`,`llamacpp_embeddings` (llama.cpp),General-purpose English embeddings with a larger context window (2048 tokens) than `bge-small-en-v1.5-f16`. `bge-m3-f16`,`llamacpp_embeddings` (llama.cpp),Multilingual embeddings with an 8192-token context window — good for long or non-English documents. `qwen3-embedding-0.6b-Q8_0`,`llamacpp_embeddings` (llama.cpp),Larger embedding model with a 16384-token context window. `qwen3-embedding-4b-Q8_0`,`llamacpp_embeddings` (llama.cpp),"Largest pre-registered embedding model (20480-token context) — higher quality, more compute." `qwen3.5-0.8b-Q8_0`,`llamacpp_generate` (llama.cpp),"Small, fast chat/completion model with a very large context window (262144 tokens)." `llama-3.2-1b-instruct-Q8_0`,`llamacpp_generate` (llama.cpp),Compact Meta Llama 3.2 instruction model for lightweight local chat/completion. `lightonocr-2-1b-Q8_0`,`llamacpp_ocr` (llama.cpp),Compact vision-language model for OCR — extracts text from images. `dummy`,`dummy`,Returns 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 :ref:`Custom models ` . BERT (``bert_local`` ) ^^^^^^^^^^^^^^^^^^^^^^ .. csv-table:: :header: Model identifier,Default,Description :widths: 12,15,25 :align: left :class: longtable `sentence-transformers/all-MiniLM-L6-v2`,OK,General-purpose English sentence embeddings; fast and small. This is what the default `bert` model uses. `sentence-transformers/paraphrase-multilingual-mpnet-base-v2`,—,Multilingual embeddings — use when your text isn't (only) English. `sentence-transformers/multi-qa-MiniLM-L6-cos-v1`,—,"Tuned for question-answering and retrieval, rather than general sentence similarity." `sentence-transformers/paraphrase-TinyBERT-L6-v2`,—,"Smaller and faster than the default, at some quality cost." .. code:: sql 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`` ) ^^^^^^^^^^^^^^^^^^^^^^ .. csv-table:: :header: Model identifier,Default,Description :widths: 12,15,25 :align: left :class: longtable `openai/clip-vit-base-patch32`,OK,The 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`` ) ^^^^^^^^^^^^^^^^^^ .. csv-table:: :header: Model identifier,Default,Description :widths: 12,15,25 :align: left :class: longtable `t5-small`,OK,"Fastest, smallest T5 variant. This is what the default `t5` model uses." `t5-base`,—,"Larger than the default — better quality, slower." `t5-large`,—,"Largest T5 variant — highest quality, most resource use." T5 variants support both text encoding (``aidb.encode_text()`` ) and text generation (``aidb.generate_text()`` ). Llama (``llama_instruct_local`` ) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ .. csv-table:: :header: Model identifier,Default,Description :widths: 12,15,25 :align: left :class: longtable `TinyLlama/TinyLlama-1.1B-Chat-v1.0`,OK,"Small, fast instruction-following chat model. This is what the default `llama` model uses." `HuggingFaceTB/SmolLM2-135M-Instruct`,—,"Extremely small and fast — lower quality, lowest resource use of these variants." `HuggingFaceTB/SmolLM2-360M-Instruct`,—,Small — a step up in quality from the 135M variant. `HuggingFaceTB/SmolLM2-1.7B-Instruct`,—,"Largest of these variants — better quality than the default, at higher resource use." All Llama variants support text generation (``aidb.generate_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. Other tested variants (llama.cpp) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Reranking (``llamacpp_reranking`` ) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ .. csv-table:: :header: Model identifier,Default,Architecture,Description :widths: 15,12,12,20 :align: left :class: longtable `gpustack/bge-reranker-v2-m3-GGUF`,—,Cross-encoder,Multilingual cross-encoder reranker. Scores each query/document pair directly via a classification head. `ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF`,—,Decoder-only,Qwen3-Reranker family. Scores each pair by reading the model's yes/no verdict-token logits — see :ref:`llamacpp_reranking` . Neither variant is pre-registered as a default model. Use either identifier with ``jsonb_build_object()`` when calling ``aidb.create_model()`` — see :ref:`llamacpp_reranking ` . Remote models ------------- Remote models are called over the network rather than run on your Postgres host: NVIDIA NIM microservices and chat/agent models from OpenAI and Anthropic — reachable through several providers, including the native ``openai_responses`` and ``anthropic_messages`` adapters. NVIDIA NIM ^^^^^^^^^^ AIDB supports NVIDIA NIM microservices via several providers — see :ref:`NVIDIA NIM ` for how to register one. Models not pre-deployed by Hybrid Manager’s Model Serving 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 `_ . OpenAI and Anthropic models ^^^^^^^^^^^^^^^^^^^^^^^^^^^ Each model identifier below can be registered under more than one AIDB provider — which one to pick depends on how you’re using the model, not which model it is: - **``openai_responses``** (+ ``openai_responses_azure`` ) and **``anthropic_messages``** (+ ``anthropic_messages_azure`` /``anthropic_messages_bedrock`` ) — AIDB’s native, agent-ready providers. Real ``tools`` /``tool_choice`` /``response_format`` support, both for direct calls and for agents created with :ref:`Managing agents ` . - **``openai_completions``** — OpenAI’s older Chat Completions API. Works with the same OpenAI models, and is a perfectly good choice for plain text generation (``aidb.generate_text()`` , summarization, pipeline steps), but rejects ``tools`` /``tool_choice`` /``response_format`` directly, and only *simulates* tool calling (via prompt injection) when backing an agent. - **``openrouter_chat``** — a single gateway to OpenAI, Anthropic, and other vendors’ models (including everything below) via OpenRouter’s own model slugs, for example ``openai/gpt-4o`` or ``anthropic/claude-opus-4-6`` . Same simulated, not native, tool calling as ``openai_completions`` . See :ref:`OpenAI Responses API ` and :ref:`Anthropic Messages API ` for the full comparison and registration examples. The tables below aren’t an exhaustive catalog of every model each vendor offers; they call out commonly used families as a starting point, covering both current flagship/efficient tiers and a longer-established variant for reference — check the vendor’s own model documentation for the full, current lineup and exact capabilities. OpenAI ^^^^^^ .. csv-table:: :header: Model identifier,Native provider(s),Also usable via,Cloud host support (native),Description :widths: 15,12,12,20,25 :align: left :class: longtable `gpt-5.6-sol` / `-terra` / `-luna`,"`openai_responses`, `openai_responses_azure`","`openai_completions`, `openrouter_chat`","Direct, Azure AI Foundry","The GPT-5.6 family, offered in three tiers trading off capability, latency, and cost. Check OpenAI's docs for how the tiers compare." `gpt-5.4-mini`,"`openai_responses`, `openai_responses_azure`","`openai_completions`, `openrouter_chat`","Direct, Azure AI Foundry","Smaller, cost-efficient tier of the GPT-5.4 generation — solid tool-calling accuracy at a fraction of flagship cost." `gpt-4o`,"`openai_responses`, `openai_responses_azure`","`openai_completions`, `openrouter_chat`","Direct, Azure AI Foundry","Long-established multimodal model. A stable, well-known anchor when you don't need the newest GPT-5.x tier." Anthropic ^^^^^^^^^ .. csv-table:: :header: Model identifier,Native provider(s),Also usable via,Cloud host support (native),Description :widths: 15,12,12,20,25 :align: left :class: longtable `claude-opus-4-6`,"`anthropic_messages`, `anthropic_messages_azure`, `anthropic_messages_bedrock`",`openrouter_chat`,"Direct, Azure AI Foundry, AWS Bedrock",Flagship Claude model for the most demanding reasoning and agentic tool-calling tasks. `claude-sonnet-5`,"`anthropic_messages`, `anthropic_messages_azure`, `anthropic_messages_bedrock`",`openrouter_chat`,"Direct, Azure AI Foundry, AWS Bedrock",Balanced mid-tier model — strong general-purpose choice for agents and everyday chat/completion. `claude-haiku-4-5`,"`anthropic_messages`, `anthropic_messages_azure`, `anthropic_messages_bedrock`",`openrouter_chat`,"Direct, Azure AI Foundry, AWS Bedrock","Fastest, most cost-efficient current Claude tier — good for high-volume or latency-sensitive tool calls." `claude-3-5-haiku`,"`anthropic_messages`, `anthropic_messages_azure`, `anthropic_messages_bedrock`",`openrouter_chat`,"Direct, Azure AI Foundry, AWS Bedrock","Long-established, widely used variant. A stable, well-known anchor when you don't need the newest Claude tier." Anthropic has no older completions-style provider the way OpenAI has ``openai_completions`` — ``anthropic_messages`` is Claude’s only direct AIDB provider, alongside the ``openrouter_chat`` gateway option. !!!note A model’s availability on Azure AI Foundry or AWS Bedrock can lag behind its release on the vendor’s own direct API, especially for very recently released tiers. If ``openai_responses_azure`` or ``anthropic_messages_bedrock`` rejects a model you expect to work, confirm current availability with the respective cloud provider.