Use LFM3-1.7B With a Signed-In Browser

LFM3-1.7B can work with a signed-in browser inside a local agent harness that can call Tabbit CLI. The model does not open pages by itself. Liquid AI's model card names Pi, Hermes Agent, and OpenClaw as agent harnesses for which it was post-trained; it lists local runtimes including llama.cpp, Ollama, LM Studio, and MLX. It also advises retrieval for knowledge-heavy work and cautions against complex multi-step reasoning. Start with a short, read-only extraction. Official model card · LFM3 release

One user building a one-call search subagent wrote that “the base lfm really surprised me at its capability to run tool calls and extract information.” Their setup was fine-tuned and narrowly scoped, so it supports trying a small task, not a broad reliability claim. Original discussion

Key takeaways

• LFM3-1.7B is the 1.73B text model, not the vision or audio variants.

• Runtime loads the model; the agent chooses tools; Tabbit CLI performs browser steps.

• Pi uses /skill:tabbit for the installed skill.

• Keep the task to a few visible fields and source links. Ask for "not visible" instead of guessing.

• Local inference does not make every website request offline.

Choose a runtime

The card lists a 256K context window and about 1 GB at 4-bit, but that does not guarantee speed on each device. For browser tasks, ask for a few exact fields rather than long reasoning over a changing workspace.

Connect Pi to a local endpoint

The model card shows a vLLM OpenAI-compatible endpoint. Pi documents custom OpenAI-compatible providers in ~/.pi/agent/models.json. Use the model name the server exposes.

vllm serve "LiquidAI/LFM3-1.7B"

{
  "providers": {
    "lfm3-local": {
      "baseUrl": "http://localhost:8000/v1",
      "api": "openai-completions",
      "apiKey": "local",
      "models": [{ "id": "LiquidAI/LFM3-1.7B" }]
    }
  }
}

This is a setup pattern, not a tested combination for every Pi or vLLM version. First confirm a normal request. Then load the discovered skill with Pi's /skill:tabbit and check the Tabbit launcher with diagnose. See the Pi model guide, Pi skill guide, and Tabbit CLI quick start.

Extract a short status digest

Read the first three visible rows on the current project status page.
Load the Tabbit skill and use the current Tabbit page only.
Return item ID, exact visible status, visible owner, and source URL.
Do not open private attachments, change filters, click actions, or edit records.
Write "not visible" for missing text. Stop if an ID is unclear.
Prepare a browser draft named "LFM3 status digest — 2026-11-09" with the table and unresolved items.
Pause at the draft. Do not save, submit, or send.

Expected result is at most three rows, exact wording, and one source link per row. The named draft is the review point. Check the title, IDs, and URLs in Tabbit before approving a write.

Decide whether the evidence passes

For browser setup, read connecting an agent to an existing browser, the first task checklist, Tabbit CLI task receipts, and CLI versus browser MCP.

Keep model limits visible

Community reports cover different setups: one user described a one-call search helper for bounded factual questions, while another tested about 50 tools inside a custom diagnostic harness. These reports do not establish a model reliability rate. Keep the task small, require source text, and test on non-sensitive content.

Use this model when local inference matters and a person can audit the short result. For long plans or many changing steps, use a stronger model or split the work into small tasks with a human checkpoint. Profiles, workspaces, and tab groups explains why login state and task organization differ; how agents control browsers gives broader setup context.

For a Codex client instead of Pi, see Codex Tabbit setup. For session behavior, see login reuse.

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.