Fun facts

A few details that don't fit anywhere else.

The world's smallest main.go

The whole binary entry point is twelve lines, of which six are imports:

package main

import (
    "context"
    "github.com/spf13/cobra"
    "github.com/ollama/ollama/cmd"
)

func main() {
    cobra.CheckErr(cmd.NewCLI().ExecuteContext(context.Background()))
}

Everything — daemon, runner subprocess, launchers, TUI — branches off cmd.NewCLI() in cmd/cmd.go.

A single binary plays four roles

ollama is its own CLI, daemon (ollama serve), inference runner (ollama runner …), and launch-integration host (ollama launch claude). The runner subprocess re-execs the same binary; see runner.Execute in runner/runner.go where the CLI dispatches between llamarunner, ollamarunner, mlxrunner, and imagegen.

A 2.8k-line route file

server/routes.go is 2,828 lines and registers every HTTP route the daemon serves — native API, OpenAI compat, Anthropic compat, image generation, web search, model recommendations. It's the largest single production file in the repo (only test files are bigger).

Two test files larger than routes.go

The largest files in the repo aren't production code at all:

1. middleware/anthropic_test.go — 3,006 lines.
2. cmd/launch/openclaw_test.go — 2,926 lines.

Both are table-driven golden-test suites; the bulk is fixture data, not assertions.

Twelve launch integrations and counting

Every file under cmd/launch/ named after a third-party tool corresponds to an ollama launch <name> integration: claude, cline, codex, copilot, droid, hermes, kimi, opencode, openclaw, pi, poolside, vscode. Each one ships with its own test file that sometimes dwarfs the implementation — openclaw_test.go (2.9k lines) tests openclaw.go (1k lines) at a 3:1 ratio.

Vendored llama.cpp

Under llama/llama.cpp/ lives a full copy of llama.cpp, kept in sync via Makefile.sync and a directory of patches in llama/patches/. The Go side wraps it through CGO in a single file: llama/llama.go. The vendor directory is large enough that the wiki's "by the numbers" stats explicitly exclude it; otherwise it would dominate every chart.

Even the deprecated parameters are documented

parser.deprecatedParameters in parser/parser.go keeps the list of parameters the Modelfile silently drops:

penalize_newline, low_vram, f16_kv, logits_all,
vocab_only, use_mlock, mirostat, mirostat_tau, mirostat_eta

Some of these were standard llama.cpp sampling options when Ollama started in mid-2023. They've stuck around in this list as a quiet record of the project's compatibility discipline.

A 5,358-commit single-branch history

git log --oneline | wc -l returns 5,358. The first commit is dated 2023-06-22 with the message "initial commit". Roughly 600 unique authors appear in git log across that span — large for a project that fits in a single Go module.