Ollama releases first version

Ollama published its first tagged release on GitHub, an early preview of a command-line tool for downloading and running open-weight language models on a local machine. The release notes were minimal — describing it only as an early preview — but the tool’s design was distinctive enough that it became, over the following two years, one of the most widely used ways for individuals and small teams to run open models without a GPU cluster or a cloud API.

Ollama’s contribution was not a new inference technique but packaging. It sat on top of llama.cpp, the C++ reimplementation of LLaMA inference that had already made running large models on consumer hardware feasible, and added a Docker-like workflow: a single command pulled a named model from a public registry, quantised for local hardware, and exposed it through a local API and chat interface. That lowered the effort required to try an open model from compiling a codebase and locating weights to typing one line, in the same way Docker had lowered the effort of running a server application a decade earlier.

The tool arrived during a period when open-weight models were shifting from research artefacts to something practitioners expected to run themselves, alongside serving infrastructure such as vLLM built for the opposite end of the market — high-throughput hosted serving rather than a single user’s laptop. Ollama’s registry format and simple interface were subsequently adopted as a de facto distribution channel by model publishers, including some frontier labs releasing open-weight variants, giving a single small open-source project outsized influence over how open models reached end users.