MiniCPM5-1B can help build a source-linked brief from public pages and a signed-in workspace when its XML tool calls are parsed by the serving runtime and a local agent can execute browser tools. OpenBMB recommends SGLang for tool use and documents the minicpm5 parser. Serve the exact openbmb/MiniCPM5-1B checkpoint, connect a local agent to Tabbit CLI, and start with read-only pages. Then save a named brief and review every source before sharing it.
OpenBMB identifies the artifact as openbmb/MiniCPM5-1B, under Apache 2.0, and its official guide specifies the SGLang parser. The Hugging Face page currently says the official repo is not deployed by an Inference Provider, so this walkthrough uses a local server.
The runtime choice is easy to get wrong. In a Reddit thread about Ollama, a user reported that a custom MiniCPM5-1B Modelfile put JSON tool arguments inside the <think> block. A reply suggested template and stop-token mismatches as a possible cause, but did not verify a fix. That is a warning about one setup, not proof that the model cannot use tools.
Key takeaways
The exact model artifact is
openbmb/MiniCPM5-1B, released by OpenBMB on May 19, 2026, under Apache 2.0.MiniCPM5-1B emits XML-style tool calls. OpenBMB recommends SGLang’s
minicpm5parser to convert them into OpenAI-compatibletool_calls.The parser, local agent, Tabbit CLI, and signed-in account each need their own check.
Separate public web research from authenticated workspace reading. Keep a URL beside each fact and flag conflicts.
Treat 1B versus 2B as a deployment choice to validate in your own harness, not as a guarantee about tool reliability.
Choose the model route and tool parser
OpenBMB’s current repository lists openbmb/MiniCPM5-1B and shows both vLLM and SGLang OpenAI-compatible server recipes. For function calling, OpenBMB specifically recommends SGLang with its minicpm5 parser. MiniCPM5 emits XML-style tool calls; the parser normalizes those into structured tool_calls for the client. A client that receives raw XML in assistant text is not ready to execute browser actions.
The documented SGLang launch pattern is:
pip install "sglang[srt]>=0.5.12"
python -m sglang.launch_server \
--model-path openbmb/MiniCPM5-1B \
--port 30000 \
--tool-call-parser minicpm5This starts a model server. It does not install an agent, provide Tabbit CLI, log into a website, or authorize access to any page. Configure your local client to use the server’s OpenAI-compatible endpoint only if the client supports a custom local endpoint and can execute local tools in that same task.
| Layer | MiniCPM5-1B setup | Verify |
|---|---|---|
| Model | openbmb/MiniCPM5-1B | Repository and checkpoint match exactly |
| Runtime | SGLang, --tool-call-parser minicpm5 | XML becomes structured tool_calls |
| Agent | Local client with the model endpoint and local tool support | It executes a harmless test function and returns its result |
| Browser | Tabbit Browser plus Tabbit CLI skill | tabbit-cli diagnose sees the browser and you confirm the signed-in account |
| Review | A separate source-linked research brief | Every claim has a URL; no source page was edited |
Keep public research separate from signed-in reading
Use public search for sources anyone can open. Use the signed-in browser only for workspace material the user has approved. Those are different evidence pools. A public result cannot confirm an internal status, and a logged-in dashboard field is not a public citation. Ask the model to keep them in different columns and include the page URL, observed value, and review flag for each item.
Before opening workspace pages, check the Tabbit CLI quick start. Supported clients may expose the skill as /tabbit; Pi uses /skill:tabbit; other clients may use /skills or a skill menu. Read the installed skill file for the actual command. Then run tabbit-cli diagnose and verify the target account and page yourself. The local task uses the existing signed-in session in that Tabbit environment; it does not pass the login to a remote runner. Skill discovery, launcher reachability, and website login are separate checks.
| Checkpoint | Ask the agent to show | Stop when |
|---|---|---|
| Public source | Page title, canonical URL, relevant excerpt | Search snippet is the only evidence |
| Signed-in source | Workspace name, page title, approved record ID, visible fields | The account, record, or permission is unclear |
| Cross-check | Which value came from which source | It merges public and private values without labels |
| Draft | Named document with source and review columns | It tries to edit, submit, or share the source |
The first-task checklist helps define read-only boundaries. Keep the model’s output narrow. A one-page brief with citations is easier to audit than a free-form report that blends public facts, internal data, and guesses.
Build a link-rich research brief
Start with three to five public pages on one question, then add one approved workspace page only if an internal decision is needed. Ask MiniCPM5-1B to capture the page title, URL, publication date if visible, exact supporting passage, and the claim it supports. For internal material, use only the fields the user named. When sources disagree, preserve both records and set needs_review: true.
Use a prompt like this:
Use the local MiniCPM5-1B agent and Tabbit CLI. First inspect these public pages: [URLs]. Then read only the signed-in workspace page [approved URL] and only the visible fields [field list]. Do not edit, share, submit, or delete anything. Create a separate draft named "Vendor research brief, 2026-11-10" with columns: source_type, title, URL, observed_fact, exact_excerpt, and needs_review. Keep public sources and workspace data in separate sections. If a page is inaccessible or two sources conflict, record the URL and stop that claim. Save the draft and show me its location for review.The names, URLs, and fields are examples. Replace them with approved sources. Do not provide passwords, session cookies, or tokens in the prompt. Review the source links in Tabbit before using the brief. A local agent can work in a separate task tab, so you can keep other work open in the main browser window. See the guide to connecting an agent to an existing browser.
MiniCPM5-1B is a compact local checkpoint, but “1B” does not tell you whether your runtime will parse every tool call correctly. OpenBMB documents the same XML-style tool-call family and parser for MiniCPM5-2B, too. The 2B checkpoint is a different artifact; do not reuse its model path or assume a larger checkpoint fixes a broken template. Test the exact checkpoint, backend, quantization, and client with a small read-only task. The Reddit report above describes a custom Ollama setup with a format problem; it does not establish that SGLang’s documented route has the same issue.
For browser mechanics, read Tabbit CLI complete guide, browser MCP versus CLI, Codex setup example, Tabbit Browser overview, browser best practices, and agentic browser basics. These explain the browser layer, not model-specific compatibility.
When to choose MiniCPM5-1B
Use 1B when your goal is a local workflow and the exact runtime fits your hardware and tool format. It is a useful candidate for extracting a short set of visible fields into a linked brief. If you need robust multi-step reasoning, long reports, or broad automation, benchmark a representative task against MiniCPM5-2B or another model before you commit. OpenBMB’s parser documentation establishes a supported format path, not an accuracy guarantee for your agent.
Keep the first run read-only. Save Vendor research brief, 2026-11-10, inspect every URL and excerpt, and only then decide whether the draft is ready to share. Download Tabbit Browser if you need the local browser session for this setup.
FAQ
What is the exact MiniCPM5-1B checkpoint?
Use OpenBMB’s `openbmb/MiniCPM5-1B` checkpoint. MiniCPM5-1B, MiniCPM5-2B, MiniCPM-V, and community fine-tunes are different model artifacts; verify the selected repository before serving.
Which runtime should I use for MiniCPM5-1B tools?
OpenBMB recommends SGLang for tool use and documents the `minicpm5` parser for the model’s XML-style tool calls. The serving layer must return parsed tool calls that your agent client can execute.
Can MiniCPM5-1B search the public web and my signed-in workspace?
It can help with both only when a local agent has tools for each source. Keep public search results separate from signed-in page data, save the source URL for each fact, and verify the workspace account yourself.
Does MiniCPM5-1B include Tabbit CLI?
No. The model produces text or tool-call output. A local agent that discovers Tabbit’s installed skill and can invoke the CLI connects those outputs to Tabbit Browser.
Why can tool calls fail in Ollama or llama.cpp?
The chat template, stop tokens, model quantization, and runtime parser can affect whether XML calls become structured tools. OpenBMB recommends SGLang for this path; test other backends separately with a read-only task.
Should I choose MiniCPM5-1B or MiniCPM5-2B?
Start with 1B when the smaller checkpoint is the constraint, but do not assume its tool-call reliability from parameter count. Both official checkpoints use the MiniCPM5 tool-call format; compare them in your exact runtime and task before choosing.