What is Ollama

What it is

Ollama is an open-source platform for running and managing large language models on your own hardware, with a hosted cloud option alongside it. It provides a command-line interface, a native graphical app, a local REST API, and model-management tools for open-weight models like Llama, Gemma, Mistral, Qwen, and DeepSeek.

The project is built by Ollama Inc., founded by Jeffrey Morgan and Michael Chiang, who previously built Docker Desktop and the startup Kitematic, which Docker acquired. The Docker lineage shows in the product: Ollama treats models the way Docker treats containers, as portable packaged units you pull and run. The first release, v0.0.1, shipped on July 8, 2023. In July 2026, TechCrunch reported a $65 million Series B led by Theory Ventures, bringing total funding to $88 million, and the company reported nearly 9 million users.

How it works

Ollama's appeal is how little friction it adds. You type a pull command to download a quantized model from its model library, then a run command to start chatting with it. Two commands stand between you and a local language model, which is why it became the default recommendation for anyone asking how to run models privately.

Under the hood it uses the llama.cpp inference backend, which is optimized for running large models on consumer hardware through quantization, the technique of shrinking model weights to lower precision so they fit in less memory. A model that needs 40GB at full precision might run in 8GB quantized, with a modest quality cost. The local server listens on port 11434 and exposes a REST API, including an OpenAI-compatible surface, so existing tools and code written against OpenAI's API can often point at a local model with a one-line change.

There is an official Docker image and client libraries for Python and JavaScript. Integrations extend it into developer workflows: it connects with Claude Code, OpenCode, and similar tools, and it can serve as the model backend for coding assistants. Experimental image generation with Flux models arrived on macOS, with other platforms pending.

Key details

The project is MIT licensed, and local use is free and unlimited. In 2026 Ollama expanded into hosted cloud models with usage-based billing for users whose hardware cannot run the models they want. The cloud tiers reported in September 2026 were a free tier with starter credits and one concurrent request, Pro at $20 per month or $200 per year with monthly usage credits and three concurrent requests, Max at $100 per month, Team at $500 per month in early access, and custom Enterprise plans. Usage credits do not roll over month to month. Those figures come from secondary sources citing Ollama's pricing page, so verify current tiers before relying on them.

Limitations

The fundamental constraint is hardware. The size of the model you can run is bounded by your machine's GPU memory, so the largest frontier models remain out of reach for local use. Quantization makes models fit, but it also costs some quality, and the tradeoff gets worse as models get smaller relative to their full-precision versions.

There is also a security footnote worth knowing. In January 2026, security researchers found many Ollama servers exposed to the public internet because the server bound to all network interfaces by default. Ollama is designed to run locally, and exposing the port publicly invites abuse, from cryptomining to serving as an open proxy. If you run it, keep it on localhost or behind authentication. The company has since tightened the defaults, but the episode is a reminder that local-first software still needs network hygiene.

Beyond the command line

Ollama's local server turned out to be as important as its CLI. The OpenAI-compatible API endpoint meant the entire ecosystem of tools built for OpenAI's API could be pointed at local models with a configuration change. Open WebUI gave users a polished ChatGPT-like interface over their local models. Coding assistants, agent frameworks, and evaluation harnesses all gained a local backend option. Ollama became infrastructure, the layer between open weights and the applications that use them.

The model library grew into a catalog of its own, with variants quantized for different hardware tiers and community-curated models alongside the official ones. The naming convention, model name plus tag for the variant, borrowed Docker's ergonomics deliberately, and it worked: pulling llama3.1:8b feels as routine as pulling a container image.

The cloud expansion in 2026 was the strategic hedge. Local inference covers experimentation and privacy-sensitive use, but many users want frontier capability without buying GPUs. The hosted tiers let Ollama serve both, with usage credits bridging the two. Whether the company can compete with dedicated inference providers on price and performance is an open question; its advantage is distribution, those 9 million users who already have the CLI installed. The $65 million Series B was a bet that the local-first developer relationship converts into a cloud business.