Skip to content

Ollama

Ollama is the person’s own server; Synoec only uses Ollama’s HTTP API and Ollama’s own app, and never installs it.

  • Hardware names this computer’s GPU and how much of its memory Ollama’s loaded models use, the processor, memory and free disk where Ollama keeps models (OLLAMA_MODELS or ~/.ollama/models). For another computer’s server, set its GPU memory there; Ollama does not report remote hardware.
  • Startup (Windows): Start with Synoec starts Ollama when Synoec opens; Start with Windows adds or removes Ollama’s own Startup-folder shortcut.
  • The server panel names the server, whether it runs, its version, what is loaded and its address. Its chevron folds the panel to that line. ⋯ switches between this computer and the workspace’s other Ollama Connections, adds one, and on this computer restarts or stops Ollama. On Windows, where Ollama quits only from its tray menu, stopping lets any reply in progress finish by unloading the models, then ends Ollama’s app.
  • Reset Ollama… in ⋯ (this computer only) lists the Ollama settings that work against Synoec, each with its value and why, and returns them to Ollama’s defaults after you confirm: the Ollama app’s context slider and network exposure (its db.sqlite on Windows), and OLLAMA_CONTEXT_LENGTH, OLLAMA_NUM_PARALLEL above 1, OLLAMA_MAX_LOADED_MODELS, OLLAMA_KEEP_ALIVE, a moved OLLAMA_HOST, OLLAMA_LLM_LIBRARY, OLLAMA_SCHED_SPREAD and OLLAMA_GPU_OVERHEAD (the user environment on Windows, launchd’s on a Mac). It then restarts Ollama. Models, OLLAMA_MODELS, flash attention and the KV cache type are kept. The Mac path is untested.
  • Installed lists the pulled models with what each can do, whether it is loaded and its size. There is no load or unload: a chat loads its model and Ollama lets it go when idle. Hovering a row shows remove; opening it shows parameters, quantization, context, maker and when it was pulled. Remove asks first.
  • Library recommends the largest model for chat, reasoning, code and images that fits the GPU, and lists every family in Synoec’s catalog with filters for tools, vision, thinking, code and embeddings. Opening a family reads each size’s download, parameters and quantization from the Ollama registry; a size it no longer has shows as unavailable. Pull any model by name takes anything ollama pull accepts. Downloads show above the list with progress and Cancel.
  • The console at the bottom takes Ollama’s words (list, ps, pull, rm, show, cp, start, quit, help) and answers each with one API request, shown as rows. It runs no shell.

The catalog is apps/desktop/src/shared/ollama-library.ts; run bun scripts/desktop/ollama-library-check.ts after editing it or before a release, which fails on a listed tag the registry lacks.