# pgvector vs Milvus: which do AI models recommend for vector databases, October 2026

IT AI Recommendation Index, October 2026 Edition, Vector databases. Eight of fourteen models named pgvector first on the direct prompt; zero named Milvus. Page: https://it-ai-index.com/it-data/vector-databases/pgvector-vs-milvus/

| | First-choice share | Rank | Negative rate | Labels | Models naming it |
|---|---|---|---|---|---|
| pgvector | 35% | #1 of 10 | 6% | 66 | 14 of 14 |
| Milvus | 3% | #6 of 10 | 20% | 41 | 14 of 14 |

## The direct prompt, model by model

- Gemini 3.5 Flash: pgvector first (first choices: pgvector) (alternatives: Pinecone, Qdrant, Weaviate)
- Perplexity Sonar: pgvector first (first choices: Qdrant, pgvector) (alternatives: Pinecone, Weaviate)
- Grok 4.1 Fast: pgvector first (first choices: pgvector) (alternatives: Pinecone, Qdrant, Weaviate)
- Mistral Small: pgvector first (first choices: pgvector) (alternatives: pgvectorscale)
- DeepSeek V4 Flash: pgvector first (first choices: Qdrant, pgvector) (alternatives: Pinecone, Weaviate)
- Llama 4 Maverick: pgvector first (first choices: pgvector)
- Kimi K2: pgvector first (first choices: pgvector) (alternatives: Pinecone, Qdrant, Weaviate)
- Muse Glimmer 30B: pgvector first (first choices: Pinecone, pgvector) (alternatives: Qdrant, Weaviate)
- GLM 4.7 FlashX: neither first, one named (first choices: Qdrant, Weaviate) (alternatives: pgvector)
- MiniMax M2.5: neither first, one named (first choices: Pinecone) (alternatives: Elasticsearch, Milvus, Weaviate)
- GPT-6 Luna: neither first, one named (first choices: Qdrant) (alternatives: Pinecone, pgvector)
- Claude Haiku 4.5: neither named (first choices: Pinecone, Weaviate)
- GPT-5.4 mini: neither named (first choices: Pinecone) (alternatives: Qdrant, Weaviate)
- Qwen 3.7 Flash: neither named (first choices: Pinecone) (alternatives: MongoDB Atlas Vector Search, Weaviate)

## What the models said about pgvector

- "| Poor filtered search | pgvector, basic setups | ... | Reindexing locks | pgvector | Large HNSW rebuilds block reads/writes" (Grok 4.1 Fast, negative prompt, soft negative)
- "pgvector when you expect very high vector-search throughput or huge scale without testing" (GPT-6 Luna, negative prompt, soft negative)
- "Be cautious at massive scale ... you must be cautious of its ceilings" (Gemini 3.5 Flash, negative prompt, soft negative)
- "Unless you have 2-3 people who can own this infrastructure, lean managed or existing DB extension first.... Already run Postgres heavily, <10M vectors | pgvector" (Kimi K2, scale prompt, first choice)
- "The best vector database for a mid-market B2B company is pgvector, which is a PostgreSQL extension. It is the strongest default for mid-market enterprises." (Llama 4 Maverick, direct prompt, first choice)
- "evaluating pgvector might save you the overhead of learning a new technology, as it handles vector search well within a relational environment" (Qwen 3.7 Flash, scale prompt, first choice)

## What the models said about Milvus

- "heavy self-hosted systems like Milvus (requires Kubernetes know-how) ... neither is friendly to a limited budget" (DeepSeek V4 Flash, budget prompt, hard negative)
- "Caution: If you run Milvus self-hosted, you must be on 2.4.24+, 2.5.21+ or 2.6.5+ and strip `sourceID` at the gateway" (Muse Glimmer 30B, negative prompt, soft negative)
- "handles billions of vectors at lower cost; requires engineering resources... more resource-intensive to operate" (Claude Haiku 4.5, comparative prompt, soft negative)
- "likely to be an open-source option such as Milvus or Chroma, as they offer free or low-cost options" (Llama 4 Maverick, budget prompt, first choice)
- "I'd recommend starting with either self-hosted Milvus/Qdrant (if you have DevOps capacity)" (Claude Haiku 4.5, budget prompt, first choice)
- "Massive Scale / Billions of vectors | Milvus" (Qwen 3.7 Flash, comparative prompt, first choice)

Share is the count of first choices across the direct, paraphrase, budget and scale prompts over all fourteen models, for a mid-market B2B company; rank is within the category. Comparisons are drawn for the top eight products in each category. Published under CC BY 4.0; the output is the models' output, and nothing here is a recommendation by the index.