ollama/ollama
Glossary
Project-specific vocabulary used throughout Ollama and this wiki.
| Term | Meaning |
|---|---|
| Modelfile | A small DSL describing a model: which base weights to use, which template, which parameters, license text, and optional adapter or projector files. Parsed by parser/parser.go. |
| GGUF | The on-disk format for quantized model weights inherited from llama.cpp. Parsed in fs/gguf/, wrapped by fs/ggml/. |
| Manifest | The OCI-ish JSON document that lists the digest of every blob (weights, adapter, template, license, params) that makes up a stored model. Defined in manifest/ and persisted under ~/.ollama/models/manifests/. |
| Blob | A single content-addressed file (usually weights) under ~/.ollama/models/blobs/sha256-.... Uploaded through /api/blobs/:digest (server/routes.go). |
| Daemon | The long-running ollama serve process. Hosts the HTTP API and the scheduler. |
| Runner | A short-lived subprocess that hosts inference for one loaded model. Spawned by the scheduler; implemented in runner/, llm/, x/mlxrunner/, and x/imagegen/. |
| runnerRef | The scheduler's handle to a loaded runner, including its model, options, and keep-alive timer (server/sched.go). |
| Scheduler | The component in server/sched.go that decides when to load, reuse, or evict runners. |
| LlamaServer | The Go interface in llm/server.go that abstracts a runner subprocess (Completion, Embedding, Tokenize, Close, …). The daemon talks to every runner through this interface. |
| llamarunner | The default runner backend; runs llama.cpp through CGO (runner/llamarunner/). |
| ollamarunner | The newer pure-Go runner; uses model definitions in model/models/ and tensor backends in ml/backend/ (runner/ollamarunner/). Activated with --ollama-engine. |
| mlxrunner | The Apple-MLX runner under x/mlxrunner/, activated with --mlx-engine. |
| imagegen | The image-generation runner under x/imagegen/. |
| Capability | A flag advertised by a model: chat, tools, embedding, vision, thinking, completion, insert. Defined in types/model/capability.go. The daemon refuses requests whose required capabilities are missing. |
| Template | The Go-style chat template stored with a model. Rendered by template/template.go before being handed to the runner. |
| Renderer | A model-specific translator that produces the final prompt for models that need more than the template (e.g., harmony-format models). See model/renderers/. |
| Parser (model) | The mirror of a renderer: parses model-specific output back into chat messages, tool calls, and thinking blocks. See model/parsers/. |
| Harmony | The structured prompt/response format used by GPT-OSS-style models. Implemented in harmony/. |
| Thinking | Reasoning content emitted by some models, gated by a think flag. Parsed in thinking/. |
| Tools | Function-calling extensions; parsed and matched in tools/. |
| Cloud model | A model name that resolves to remote inference on ollama.com rather than a local runner. Routed through server/cloud_proxy.go. |
| Launch integration | A third-party client (Claude Code, Codex, Droid, etc.) that ollama launch configures to use Ollama as a backend. Implemented under cmd/launch/. |
| Selector / TUI | The bubbletea-based pickers used by ollama launch and friends. Defined in cmd/tui/. |
| Background app | The macOS / Windows tray app under app/ that keeps the daemon running and surfaces UI affordances. |
| Sign-in | Linking a local installation to an Ollama account by uploading the public half of an SSH key. Handled by signinURL, WhoamiHandler, and SignoutHandler in server/routes.go. |
| Keep-alive | How long the scheduler keeps a runner loaded after the last request. Per-request via the keep_alive field; default in OLLAMA_KEEP_ALIVE. |
| Adapter | A LoRA or similar fine-tuning weight stack referenced by ADAPTER in a Modelfile. |
| Projector | A vision encoder paired with a language model; referenced as a separate blob and loaded by the runner for multimodal models. |
| OpenClaw | The Ollama-published WhatsApp/Telegram/Slack/Discord assistant; one of the launch integrations (cmd/launch/openclaw.go). |
Built by Factory AutoWiki from public repository content. It is a generated preview for codebase exploration, not source-maintained documentation.