MiniCPM5-2B can support a signed-in browser workflow if its serving runtime parses the model’s XML tool calls and a local agent maps them to Tabbit CLI. The model does not open your dashboard by itself. Start with OpenBMB’s recommended openbmb/MiniCPM5-2B checkpoint, configure SGLang with the minicpm5 parser, and verify a harmless read-only call before asking the agent to compare dashboard changes. OpenBMB explicitly documents this model and parser pairing in its MiniCPM5-2B guide.
The exact checkpoint and runtime matter. One Reddit tester said MiniCPM5 2B Q4_K_M had disappointing tool calling in their benchmark. A reply said Q8 worked with OpenCode for repo exploration and web search in that person’s setup. A separate release discussion includes users describing small-model agent work and browser tasks. These are individual reports, not a reproducible Tabbit benchmark. They point to the setup checks this guide makes explicit: weights, quantization, parser, tools, and review step.
Key takeaways
The documented checkpoint is
openbmb/MiniCPM5-2B; OpenBMB currently recommends BF16, GGUF, MLX, or GPTQ deployment formats.Tool calling requires a runtime that understands the model’s XML format. OpenBMB recommends SGLang’s
minicpm5parser.Tabbit CLI belongs to the local agent layer. The model’s tool-call output does not directly reach your browser without a compatible agent and runtime.
Use a read-only dashboard task first. Compare fields by stable ID, include source URLs, and stop on missing or conflicting values.
Save a named draft with a visible review checkpoint. Keep the agent away from submit or update controls until you approve the change.
MiniCPM5-2B and browser task roles
| Layer | What to configure | Verify before continuing |
|---|---|---|
| Model checkpoint | openbmb/MiniCPM5-2B with a deployment format supported by your runner | Do not confuse it with Base, SFT, Midtrain, MiniCPM5-1B, or a community fine-tune |
| Serving runtime | A runtime that supports MiniCPM5’s chat template and XML tool calls | For SGLang, OpenBMB documents --tool-call-parser minicpm5 |
| Local agent | A client that can select the local endpoint and execute local tools | Confirm the client discovers Tabbit’s skill and supports its CLI calls |
| Browser | Tabbit Browser with the intended site already signed in | Check the tab URL, account, and read-only scope yourself |
| Human review | A named report or draft | Verify every changed field before any write action |
Run the exact checkpoint before connecting an account
OpenBMB’s current guide recommends MiniCPM5-2B in BF16, GGUF, MLX, or GPTQ form. It lists Base, Midtrain, SFT, and DSpark checkpoints separately because they serve different purposes. Begin with the recommended checkpoint at openbmb/MiniCPM5-2B, not an unofficial fine-tune or another MiniCPM family member. For SGLang, the official guide shows an OpenAI-compatible endpoint and recommends the built-in minicpm5 parser for tool use. The model emits XML tool calls; the parser converts them to OpenAI-compatible tool_calls fields. Read the official model guide and Hugging Face model card for the current deployment details.
A local agent may let you select an OpenAI-compatible endpoint, but that does not automatically prove it can use MiniCPM5’s tool format. Confirm that your client parses the tool_calls responses and exposes local tools. The public docs do not establish that every agent client supports this exact model configuration. If your setup cannot run tool calls, use MiniCPM5-2B only to plan or summarize text, then hand a structured task list to a local agent or perform the browser steps yourself.
Set up Tabbit independently of model hosting
Install Tabbit Browser and sign into the dashboard in the profile you want to use. In the local agent client, check that the Tabbit skill is discoverable before invoking it. Tabbit’s CLI quick start shows /tabbit for supported skill-aware clients. Pi documents /skill:tabbit, and some clients expose skills through /skills or a menu. Read the installed SKILL.md rather than guessing which command applies to your client.
Then check launcher access with tabbit-cli diagnose from the same machine. Finally, look at the target page in Tabbit and confirm the correct dashboard account is signed in. These are separate gates: model endpoint selection, skill discovery, local launcher reachability, and website login. A remote model endpoint can produce a response while the local agent still has no connection to Tabbit. Conversely, a connected browser does not prove the model endpoint can emit usable tool calls.
| Gate | Check | Stop if |
|---|---|---|
| Exact model | The client selects openbmb/MiniCPM5-2B on the intended local endpoint | It silently selects another checkpoint or provider |
| Tool format | The server parses MiniCPM5 XML into tool calls | Output is malformed, wrapped as plain chat, or loses required arguments |
| Skill discovery | The agent can see Tabbit’s installed skill | The chosen skill command is unknown to that client |
| Browser connection | tabbit-cli diagnose sees the running controller | The launcher cannot reach Tabbit on this device |
| Signed-in page | You confirm the dashboard account and exact route in Tabbit | The account, record, or page is not the intended target |
A local Tabbit CLI fixture test in this environment confirmed session reuse between two named tasks. It did not pair MiniCPM5-2B with a client or test the XML parser. Treat the browser connection and model integration as two independent checks.
Use a dashboard change report as the first recipe
Pick one dashboard and one time window. Ask the local agent to read only visible status fields for a stable record ID and compare them with the previous approved snapshot. The output should show current value, prior value, source URL, and review status. Do not let a small model infer business impact or update the dashboard. If a field is missing or a record ID changes, stop and ask for review.
A practical workflow looks like this:
Start the local model server and confirm the exact checkpoint in the agent’s model selector.
Run a harmless tool-call test that returns a fixed value. Confirm the runtime parsed it as a tool call, not assistant prose.
Start the Tabbit task. Ask the agent to open the dashboard and report only the page title and account label.
Read one stable record ID and return the relevant fields. Do not click update or submit controls.
Compare with the prior snapshot and save a named draft such as
Weekly dashboard changes, 2026-11-09.Review the draft in the browser or editor. Only then decide whether a human should make a change.
Use a prompt that forces the model to stop when the record identity is uncertain:
Use Tabbit CLI to read the dashboard page already open for this task. Read only record ID GEO-104 and the fields status, owner, and updated_at. Do not click edit, save, submit, or delete. Return the page title and URL, record ID, each visible field value, and a needs_review flag. If the ID or any field is missing, stop and explain what you could not verify. Compare only against the prior snapshot I provide. Save a draft named "Weekly dashboard changes, 2026-11-09" and stop before any update.GEO-104, the field names, and the date above are test parameters. Replace them with a harmless fixture or an approved record. Do not give the agent permission to update a live dashboard during the initial test.
Keep the report auditable
Make each output row link back to its page. A reviewer should not need to search again to verify the current status. Use separate columns for the current value, prior value, source, and decision. If values conflict, preserve both observations and flag the difference. Do not ask MiniCPM5-2B to fill gaps from memory.
The browser-agent first-task checklist covers read-only boundaries. The Tabbit CLI complete guide and agent connection guide explain the browser layer. To compare connection mechanisms, see browser MCP versus CLI. For a client example, read Codex with Tabbit CLI. The Tabbit Browser overview, its best practices, and the agentic browser explainer explain the product and its limits.
Community results for MiniCPM5-2B are mixed across quantization and runner. One tester reported weak tool calling with Q4_K_M, while another said Q8 worked adequately for repo exploration and web search in OpenCode. OpenBMB’s own guidance specifies the parser used for its XML tool calls. Use the official runtime configuration, then test your own dashboard fields with a read-only fixture before you rely on it.
For a compact local dashboard check, MiniCPM5-2B is a model candidate, not a complete browser automation stack. The runtime must parse calls, the agent must be able to run Tabbit CLI locally, and a human should approve the named draft. Download Tabbit Browser if you need a local browser session for that workflow.
FAQ
Which MiniCPM5-2B checkpoint should I use?
OpenBMB currently recommends MiniCPM5-2B in BF16, GGUF, MLX, or GPTQ formats for deployment. Base, Midtrain, and SFT are different training stages, so do not substitute them without checking their model cards.
Can MiniCPM5-2B call browser tools?
The model can emit XML-formatted tool calls, but a serving runtime must parse them. OpenBMB recommends SGLang with the `minicpm5` tool-call parser; the agent still needs a browser tool such as Tabbit CLI.
Does MiniCPM5-2B run the browser on its own?
No. The model generates text or tool-call output. A local agent and its runtime connect that output to Tabbit CLI, which controls the browser tabs.
Can it read my dashboard while I am signed in?
A local agent using Tabbit CLI can read a dashboard in the Tabbit profile where you signed in, once you verify the site and permissions. The model itself does not inherit a browser login.
What if the tool call is malformed?
Stop before any write action. Check the exact checkpoint, chat template, runtime parser, tool schema, and agent support, then repeat a harmless read-only test.
Can I use a remote MiniCPM endpoint with Tabbit CLI?
Only if the remote agent has a verified connection to the same machine and Tabbit browser. Otherwise, use a local handoff: export approved page facts for the remote model or run the model and browser agent locally.