provider

Ollama (local models) with Empryo

Run open-weight models entirely on your own machine and point Empryo at them. Ollama is one of Empryo's 21 built-in providers, auto-detected at `localhost:11434` with no API key and no cloud round-trips. Pair a local model with Empryo's Genome graph and AST editor for fully offline, zero-cost agentic coding.

Popular models

qwen3-coder (30b, 480b) — tool-calling, long-context codingqwen3 (0.6b–235b) — general, with tools + thinking variantsgpt-oss (20b, 120b) — OpenAI open-weight, reasoning + toolsdevstral (24b) — tuned for coding agentsdeepseek-r1 (1.5b–671b) — open reasoning, tools + thinkingdeepseek-v3 (671b) — large MoEllama3.3 (70b) — tool-calling general modelgemma3 (270m–27b) — single-GPU, visionmistral-small (22b, 24b) — small-tier tool-callingqwen2.5-coder (0.5b–32b) — lightweight code models

Setup

1. Install and run Ollama

Install Ollama from ollama.com, then pull a model with strong tool-calling (Empryo's agent loop relies on it):

ollama pull qwen3-coder:30b

Ollama serves an OpenAI-compatible API at http://localhost:11434 by default.

2. No API key needed

Ollama is a built-in provider in Empryo. It is a local, auth-less server, so there is nothing to set with --set-key. Empryo auto-detects it at localhost:11434 and lists any models you have pulled.

3. Point Empryo at a remote or custom host (optional)

If Ollama runs on another machine or port, set the OLLAMA_HOST environment variable before launching:

OLLAMA_HOST=http://192.168.1.50:11434 empryo

4. Pick the model

Launch Empryo and press Ctrl+L to open the model picker, then choose your Ollama model. To verify detection from the shell:

empryo --list-providers          # confirm Ollama shows up
empryo --list-models ollama      # list your pulled models

You can also pin it as the default in ~/.empryo/config.json:

{
  "defaultModel": "ollama/qwen3-coder:30b"
}

5. Mix local and cloud with the task router (optional)

Empryo's per-slot task router lets you run a small local model for cheap, high-frequency jobs while keeping a stronger model (local or cloud) for code writing:

{
  "taskRouter": {
    "spark":   "ollama/qwen2.5-coder:7b",
    "semantic":"ollama/qwen2.5-coder:7b",
    "ember":   "ollama/qwen3-coder:30b"
  }
}

When to use it

Pick Ollama when privacy, offline operation, or zero per-token cost matter more than raw frontier capability: regulated codebases, air-gapped machines, or heavy experimentation where cloud bills would add up. Because everything runs on your hardware, nothing leaves the machine, and Empryo's local-first features stack well here — the Genome graph (PageRank + git co-change ranking) and the AST/LSP editor that edits "symbols, not strings" both run locally too, so the only network calls a cloud setup would make disappear entirely. The honest tradeoff: local open-weight models are weaker than frontier cloud models at complex multi-step agentic work, and the larger, more capable options (qwen3-coder:480b, deepseek-v3:671b) need serious VRAM. For day-to-day local use, tool-calling models in the 24–32B range (devstral:24b, qwen3-coder:30b, gpt-oss:20b) are the sweet spot, and Empryo's agent loop depends on solid tool-calling, so prefer models tagged for tools. If you want frontier quality, use a cloud provider for the `ember` (code-writing) slot and keep Ollama for cheaper research and summary slots via the task router.

Get Empryo All providers