Ollama

Ollama exposes an OpenAI-compatible chat completions API at http://localhost:11434/v1 by default. summarize talks to it directly — no API key, no cloud round-trip, no data leaves the machine.

#Quick start

# 1. Pull a model that fits your VRAM
ollama pull qwen3:14b

# 2. Run summarize against it
summarize "https://example.com" --model ollama/qwen3:14b

That's it. No env vars are required for the default http://localhost:11434/v1 endpoint.

#Configuration

#CLI

summarize <url> --model ollama/<model>

The <model> part must match the tag in ollama list exactly, including the variant suffix:

• ollama/qwen3:14b
• ollama/llama3.1:8b
• ollama/gemma3:12b-it-q8_0

#Config file (~/.summarize/config.json)

{
  "model": "ollama/qwen3:14b",
  "ollama": { "baseUrl": "http://localhost:11434/v1" }
}

ollama.baseUrl is only needed when pointing at a remote Ollama instance or a non-default port.

#Environment

OLLAMA_BASE_URL also gates auto-discovery in the Chrome/Firefox extension model picker — set it (or ollama.baseUrl in config) and your installed Ollama models appear in the dropdown automatically.

#Remote Ollama

If Ollama is running on another machine on your LAN (or behind Tailscale):

export OLLAMA_BASE_URL=http://gpu-rig.lan:11434/v1
summarize "https://example.com" --model ollama/qwen3:14b

#Auth-fronted Ollama

If you've put an auth proxy in front of Ollama, set OPENAI_API_KEY — summarize will forward it as the Authorization: Bearer … header. Bare Ollama ignores the header (any value works), so a dummy is also fine.

#Model recommendations

For summarization quality on a 16 GB consumer GPU at Q4KM quantization (~10 GB on disk):

• qwen3:14b — strong instruction-following, ~128K context, current generation as of 2026.

Excellent for long articles, podcasts, and YouTube transcripts.

• gemma3:12b — newer Gemma generation, instruction-tuned, lighter VRAM footprint, leaves

headroom for longer contexts.

• mistral-small:24b — biggest that comfortably fits in 16 GB at Q4; better quality but slower

TTFT and tighter VRAM headroom.

Smaller models (≤8B parameters) tend to hallucinate specifics (model numbers, dates, names) on long-form content. If you summarize YouTube transcripts or long articles, prefer a 12B+ model.

#Limitations

• No document attachments. Ollama models don't accept PDFs; --format markdown with

--markdown-mode llm still works (HTML in, markdown out), but binary .pdf inputs require an Anthropic/OpenAI/Google model that supports document attachments.

• No native video understanding. Use the transcript path (default for YouTube) instead of

--video-mode understand.

• Quality is bounded by the local model. Don't expect GPT-5-class output from a 14B local

model. Run an A/B against the same input to calibrate expectations.

#Transport details

Ollama is invoked via its OpenAI-compatible /v1/chat/completions endpoint. summarize forces chat-completions mode (the OpenAI Responses API isn't supported by Ollama) and sends the model id verbatim (qwen3:14b, not openai/qwen3:14b).