More and more companies prefer to run AI in-house — for privacy, cost and control. We count the daily downloads of the key tools that make it possible, turning the local-AI adoption race into something you can watch.
Cumulative downloads of the most-used self-hosted AI images on Docker Hub — from model runtime to vector database. Sorted by downloads; the columns show the gain.
We start recording the daily gain today — the gain columns fill in from tomorrow. Today you see the current standing.
| Tool | Downloads | 24 h | 7 days | 30 days |
|---|---|---|---|---|
| Model runtime | ||||
Ollamaollama/ollama | 149.9M | – | – | – |
vLLMvllm/vllm-openai | 26.4M | – | – | – |
LocalAIlocalai/localai | 4.8M | – | – | – |
LiteLLMlitellm/litellm | 774.4K | – | – | – |
Text-Gen-WebUIatinoda/text-generation-webui | 252.5K | – | – | – |
| Chat & apps | ||||
Difylanggenius/dify-api | 20.7M | – | – | – |
Flowiseflowiseai/flowise | 6.6M | – | – | – |
LobeChatlobehub/lobe-chat | 5.8M | – | – | – |
AnythingLLMmintplexlabs/anythingllm | 3.6M | – | – | – |
| RAG & vector DB | ||||
Milvusmilvusdb/milvus | 77.2M | – | – | – |
Qdrantqdrant/qdrant | 35.9M | – | – | – |
Weaviatesemitechnologies/weaviate | 18.9M | – | – | – |
Chromachromadb/chroma | 6.1M | – | – | – |
RAGFlowinfiniflow/ragflow | 3.3M | – | – | – |
| Orchestration & ops | ||||
n8nn8nio/n8n | 229.1M | – | – | – |
Langfuselangfuse/langfuse | 11.9M | – | – | – |
Which open models are people actually downloading to run locally? The top models in the Ollama library by downloads. Figures rounded (~) as reported by Ollama.
llama3.1~116.7Mdeepseek-r1~89Mnomic-embed-text~77Mllama3.2~75Mgemma3~38.3Mqwen2.5~34Mqwen3~31.7Mmistral~30.7Mgemma2~26.9Mllama3~24.6Mqwen2.5-coder~18.1Mphi3~17.8Mgemma4~16.7Mqwen3.5~14.6Mllava~14.3Mmxbai-embed-large~12.1Mgpt-oss~10.7Mphi4~7.6Mgemma~7.2Mllama2~7.2Mqwen~7.1Mqwen3-coder~6.9Mglm-ocr~6Mqwen2~6Mcodellama~5.7Mminicpm-v~5.3Mmistral-nemo~5.2Mtinyllama~5.2Mbge-m3~5Mllama3.2-vision~4.8MHonest and reproducible — we only read public, official counters.
For each image we read the public, exact total download count (pull_count) straight from Docker Hub — no key, no login.
From ollama.com we take the reported per-model downloads. These are rounded (e.g. "116.7M") — we mark them with ~.
A job records one value per day. From the daily values we derive the gains (24 h / 7 days / 30 days).
Neither source keeps a history. Our series can only be built forward and grows from the start — nobody can reconstruct it after the fact.
"Downloads" means how often an image or model was pulled — a proxy for adoption, not active use. CI systems and automated deployments count too. Official images that live only on ghcr.io (e.g. Open WebUI) can't be tracked, because there is no public download figure there.
AI models you run on your own hardware or in your own cloud instead of via a third-party API. The upside: your data never leaves the building — a strong argument for privacy and GDPR.
Downloads are the only public, reliable adoption metric that Docker Hub and Ollama expose. They track growth and trend well; they don't directly measure active use.
Neither Docker Hub nor Ollama stores a history — they only ever show the current total, overwritten daily. Anyone who starts counting later has missed the past days forever. That's exactly what makes the series valuable.
A curated set of the most common self-hosted AI tools: model runtimes (Ollama, vLLM, LocalAI), chat UIs and app builders (LobeChat, AnythingLLM, Dify, Flowise), RAG and vector databases (Qdrant, Chroma, Weaviate, Milvus) and orchestration (n8n, Langfuse).
Sources: Docker Hub (hub.docker.com) and the Ollama library (ollama.com), both publicly available. Download counts are a proxy for adoption. No warranty for completeness or timeliness of individual values.
We bring self-hosted AI to your infrastructure — privacy-compliant, maintainable and productive.