Use Gemma 4 31B IT for browser research with a local agent

Gemma 4 31B IT can support browser research through a local agent, provided the serving stack handles its structured tool calls. Keep public-source research separate from signed-in work: collect public release notes first, then inspect an internal page only when you have permission and need that source. Tabbit Browser and Tabbit CLI can give the local agent a separate task tab in the browser environment where you signed in. The model does not authenticate you or grant access.

Key takeaways

• The official checkpoint is google/gemma-4-31B-it, an instruction-tuned, dense Gemma 4 variant.

• Google documents structured function calling; your inference server and agent client still need a compatible template, parser, and executor.

• Public pages need no signed-in session. Add private pages only when the account is authorized and the task requires them.

• Keep the task read-only and save a named brief with exact URLs. A person checks it before sharing or acting.

• Google lists 31B for large servers or clusters. Check hardware and exact provider availability before choosing where to serve it.

Set a source boundary before opening the browser

A release brief may combine public notes with one internal status page. These are different source classes. Give the agent an allowlist: the public release URL, the one approved workspace URL, and the fields to read. If the private page redirects to sign-in, denies access, or asks for a higher role, stop there. Do not ask the model to find another route.

Google’s model card identifies google/gemma-4-31B-it, instruction tuning, and Apache 2.0 licensing. Its function-calling guide shows declared tools. Hugging Face documents vLLM and SGLang serving examples for this exact checkpoint. Google’s Gemini API docs also show Gemma 4 model IDs and function declarations. These sources establish available documented paths, not universal support across every host or agent. E2B and E4B are separate smaller variants, not names for this 31B model. Check current host support and run a harmless tool call before connecting a real account.

Verify each part of the local setup

Check the exact model ID and revision, license, provider permissions, context limit, and function-calling support in the runtime you plan to use. An API model listing is evidence for that API only. It does not prove that a third-party client, another region, or a local server exposes the same route.

Then check how your agent client discovers the Tabbit skill. Pi documents /skill:tabbit; other clients may have a /skills picker or automatic discovery. Follow that client’s instructions; do not assume /tabbit works everywhere. Separately run the installed Tabbit CLI diagnostic to verify launcher reachability. Skill discovery does not prove the CLI is installed, and a working CLI does not prove that the target site is signed in.

Install Tabbit Browser and sign in yourself in the intended profile. Start a named local task. Its task tab can use that browser environment while you continue working in your other tabs. A remote agent needs a separately supported local handoff; the remote model cannot reach a laptop browser merely because the site is signed in there. See the Tabbit CLI setup and connect a local agent.

Produce a source-linked brief

Use a public release page and, only if authorized, one internal status page. Limit collection to the release ID, status, and displayed date. This keeps unrelated workspace data out of the brief.

Use the local Tabbit browser through the installed Tabbit CLI.
Task name: GEMMA-REL-2026-11-10 source brief
Read the public release notes at [public URL]. Only if my signed-in account is authorized, open [approved internal URL] and read the visible release ID and status.
Save a draft named "Gemma release source brief - review required". Include a table with source type, release ID, status or finding, displayed date, and exact URL. Separate public claims from internal observations. For missing fields write "not visible". Do not infer dates, copy personal data, edit, reply, submit, or change settings. Stop if sign-in or additional permission is required.
End with three checks a human should make before sharing.

If the two pages disagree, show both values with their URLs. Do not combine them into a single statement. The [first-task checklist](first-task checklist) and [permission guide](CLI or MCP) help define safe stop points.

A practical fit for Gemma 4 31B IT

Use this model when your serving stack supports the exact checkpoint and your hardware can run the dense 31B variant. The model proposes a function call; the host validates and executes it. Tabbit CLI is the local browser handoff, not a browser feature inside the model. If the agent client cannot reach a local Tabbit instance, have it prepare a checklist for a supported local handoff instead.

Two Reddit users describe tool-call friction. One asks whether Gemma variants return the structured format their internal tools need. Another says a 26B agent produced notify:notify instead of the expected tool name. Those are individual reports, not a failure rate, but they justify testing your exact template and parser. Read the tool-use question and function-name report.

Before sharing the brief, open every cited URL and verify the release ID, date, and field labels. The signed-in session guide covers local handoff choices; the browser profiles and task tabs explains signed-in browser sessions. For additional checks, see local web app checks, permission boundaries, and Codex CLI setup.

The current vLLM Gemma 4 recipe shows this family using the gemma4 parser, --enable-auto-tool-choice, and a Gemma 4 chat template. This is a documented serving path, not proof that your agent client uses it. vLLM documents Gemma 4 parser fallbacks, including non-standard call forms and response-tag variation; test actual output with a harmless function.

vllm serve google/gemma-4-31B-it \
  --host 127.0.0.1 \
  --max-model-len 8192 \
  --enable-auto-tool-choice \
  --tool-call-parser gemma4 \
  --reasoning-parser gemma4 \
  --chat-template examples/tool_chat_template_gemma4.jinja

vLLM Gemma 4 chat template

vLLM lists one 80 GB NVIDIA GPU as its BF16 baseline for dense 31B. Treat this as a documented reference configuration, not a guarantee for every quantization or hosted service.

Review decision points

Official model references

Official sources: 31B IT model card, function-calling guide, Gemma on Gemini API, and deployment guide.

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.