LFM2.5-24B-A2B can work with a signed-in browser when a local agent can select the model, run its tools on your computer, and invoke Tabbit CLI. Liquid AI publishes the checkpoint and documents Pi among the agent harnesses used in post-training. That is a reasonable place to start, but the model, Pi, local inference server, Tabbit CLI, and your website login are separate pieces. A Reddit user testing the model in Pi reported strong speed and also mentioned early tool-call template failures, so validate the exact runtime before trusting a task. Liquid AI release · Hugging Face model card · original Pi discussion
Key takeaways
The checkpoint is open-weight, text-only, and published with a 131,072-token context window. The post-trained variant is intended for agentic and tool-use work.
The model card documents Pi and several local runtimes, but that does not guarantee your Pi build, quantization, or server handles tool calls correctly.
Pi discovers a Tabbit skill by name and invokes it with
/skill:tabbit. Skill discovery does not prove the CLI launcher is reachable or that the selected site is signed in.Sign in to the target site yourself in the intended Tabbit profile. A new task can use that browser environment, while its task tab keeps the work separate.
Begin with a read-only, single-record check. Stop at a named draft and compare every field with the source page.
Choose a route that fits your machine
| Route | What the official sources confirm | Decision |
|---|---|---|
| Local checkpoint with Pi | Liquid AI lists Pi as an agent harness used during post-training; the model card links local runtime and Pi setup material. | Use if your hardware and selected quantization can serve the model and the local endpoint passes a tool-call check. |
| Liquid Console API | Liquid AI lists API inference and current token pricing on its release page. | Use only when hosted inference is approved for the page data. It is not local inference. |
| Another host or client | This article did not verify a working pairing. | Check the exact model ID, provider, tool-call format, and host shell access first. |
The checkpoint has 24B total parameters and 2.45B active parameters according to its card. Liquid AI reports a 131,072-token context window. Those specifications do not tell you how much memory your chosen quantization needs. The model card also says it is not recommended for knowledge-heavy work without retrieval or web search. For a browser task, use the page itself as the source rather than asking the model to answer from memory. Model card: details and tool use
Connect Pi to Tabbit CLI on the same computer
Install Tabbit Browser and follow the CLI setup guide. In Pi, confirm the model appears in /model and can complete a harmless function-call task against your local runtime. Then check that Pi has discovered the installed tabbit skill. Use /skill:tabbit to load it. Pi's exact skill command is different from Codex's /skills picker and OpenCode's @tabbit syntax; use the syntax documented by your client. Pi skills documentation · Pi model configuration
Run the installed Tabbit launcher’s diagnose. If it cannot reach the expected Tabbit instance, stop there. If it can, open the website in that instance and sign in yourself. Skill discovery, launcher reachability, and website authentication are three separate checks. The existing-browser connection guide and session reuse guide explain what a selected browser profile can carry over. A site may still expire a session or ask for another sign-in. The CLI and MCP comparison helps if your host cannot run a local command.
Test one record without changing it
Choose a harmless page with one stable record ID. The prompt below asks the agent to read only visible information and stop before any write.
Use the Tabbit skill to inspect the currently open [approved dashboard] in the signed-in Tabbit profile.
Read only the visible row for record [ID]. Do not open attachments, hidden fields, or other accounts.
Do not edit, assign, save, submit, or send anything.
Return the exact record ID, visible status, title, and page URL. Write "not visible" for missing fields.
Prepare a browser draft named "LFM2.5 record check — [date]" with the values and any uncertainty.
Stop before saving or sending. Show me the draft and the source URL for review.A useful output has one row, the same ID you supplied, a source URL, and literal values from the page. The model should not fill gaps from its training data. If Pi returns malformed tool calls, recheck the model template and serving runtime before changing the prompt. The model card explains its tool-call format; the local model community also reported template fixes after release. Treat those reports as individual runtime experiences, not a general failure rate. Local model discussion
Keep the review checkpoint visible
| Checkpoint | What you verify | Stop condition |
|---|---|---|
| Model | Pi selected the intended checkpoint and provider. | The displayed name or endpoint differs from your test setup. |
| Tool loop | A harmless local tool call returns valid output. | Arguments are malformed, repeated, or missing. |
| Browser | Account, workspace, record ID, and URL match. | Any of them points to the wrong tenant or user. |
| Draft | Every value matches the visible page. | Any value is inferred or cannot be traced. |
Keep editing and sending as separate approvals. A successful read does not authorize a write. The first-task checklist covers a safe initial run; the Codex setup article shows another local-client entry point. To understand task separation, see profiles and tab groups and how agents control browsers. The complete CLI guide has the setup details.
LFM2.5-24B-A2B is worth trying for a local, tool-using workflow when your hardware can serve it and the exact runtime passes a tool-call check. It is a poor choice if you need a verified out-of-the-box Tabbit integration or cannot validate your local template. Use a read-only task first, review the named draft, then decide whether any specific change should be applied.
FAQ
Can LFM2.5-24B-A2B use my signed-in browser?
Yes, when a local agent runs Tabbit CLI in the Tabbit profile you signed into. The model does not connect to the browser by itself.
Does Liquid AI guarantee Pi and Tabbit compatibility?
No. Liquid AI documents Pi as an agent harness used in post-training, but you still need to verify your local model server, tool-call template, and Tabbit CLI connection.
What is the Pi command for the Tabbit skill?
Pi documents named skills with the `/skill:<name>` command, so the Tabbit skill is `/skill:tabbit`. Confirm the skill is installed and discovered first.
Can this model access current page facts from memory?
Do not rely on model memory for current dashboard values. Ask the agent to read the visible page and return the exact source URL.
Should the agent change records after reading them?
Not by default. Start read-only, inspect the named draft, and approve any specific write separately.