To inspect a signed-in preview with DiffusionGemma 26B A4B IT, install Tabbit Browser and connect a compatible local agent through Tabbit CLI. Verify the exact diffusion checkpoint, serving build and tool parser before the browser step.
Key takeaways
DiffusionGemma uses discrete diffusion. Choose a serving build that explicitly supports this checkpoint and verify the tool-call path.
A model response is not a browser action. The local agent must expose tools, the runner must parse calls, and Tabbit CLI must reach the intended browser.
Check model and runtime versions, parser support, and a harmless tool-call fixture before opening a real account.
Tabbit Browser can keep the existing site login in its target environment while a task uses a separate tab.
Treat page text as untrusted input; pause before submitting, publishing, buying, or changing account settings.
| Gate | Verify | Pass evidence |
|---|---|---|
| Checkpoint | Exact Google DiffusionGemma revision and license | Match the official model card |
| Runtime | Diffusion-aware architecture support | Current version docs list this checkpoint |
| Serving config | Parser, chat template, diffusion flags, memory limits | Minimal text request succeeds |
| Tool path | Structured call parsed and executed once | Harmless function fixture passes |
| Browser handoff | Local agent can run Tabbit CLI | Intended Tabbit instance and task tab are visible |
| Review | Read-only note saved | Human checks note before any write action |
Confirm DiffusionGemma’s serving path first
DiffusionGemma is not a drop-in Gemma 4 decoder. Google’s official model card describes a discrete-diffusion, block-generating model based on the Gemma 4 26B A4B MoE. Its inference path differs from token-by-token autoregressive generation. The card lists the Apache 2.0 license and Transformers support.
Before choosing a host, confirm its current catalog contains the exact checkpoint. For local serving, the vLLM recipe specifies diffusion-aware support, a Gemma 4 tool parser and chat template, and model-specific options. Flags and hardware requirements change by build, so pin and read the version you intend to run. Do not paste a recipe into a different server and assume it works.
Then test the full path with a harmless declared tool. Confirm the model emits a call, the parser recognizes it, and the host executes it exactly once. Only after that should a local agent hand browser work to Tabbit CLI. A tool call in a model response is not evidence that a browser action happened.
Worked example: inspect a preview page and save a review note
Use a separate Tabbit task tab to read the visible title and status on a signed-in preview page. Save the result as diffusion-preview-review-104.md, with the page URL, title, status, and any mismatch. Do not edit or submit the page. This is a proposed workflow; it is not a benchmark or a claim that every DiffusionGemma runtime works with a browser agent.
Use this prompt:
Read the visible title and status from the signed-in preview page in this Tabbit task tab. Do not click submit, publish, save, or account controls. Save the URL, title, status, and any mismatch in `diffusion-preview-review-104.md`. If the page or status is unclear, stop and say what needs human review. Do not copy personal data.| Job | Owner | Stop for review |
|---|---|---|
| Check model availability | Runtime/provider operator | Exact model ID and version |
| Validate parser and flags | Local inference operator | Harmless tool-call fixture |
| Read signed-in preview | Local agent through Tabbit CLI | Named review note saved |
| Revise copy | DiffusionGemma or another approved model | Human edits and approves draft |
| Publish or change state | Human | Confirm exact page and action |
Decision: validate the serving stack before involving the browser
DiffusionGemma can be used in a browser workflow only when the serving stack supports its discrete-diffusion generation path and the required tool-call parser. A model card or model name alone does not establish that. Verify a precise runtime build with a harmless fixture, then connect a local agent to Tabbit Browser using Tabbit CLI. If the host cannot execute tool calls reliably, keep DiffusionGemma on drafting and summarization and let another verified local agent perform browser actions.
Continue with the browser workflow
Tabbit CLI setup · connect a local agent · first browser-agent task · CLI versus browser MCP · Codex setup · reuse a signed-in session · browser session basics · local web app checks · permission boundaries
FAQ
Does DiffusionGemma support tool calling?
Google’s model card identifies DiffusionGemma as a distinct discrete-diffusion model. A compatible serving stack must implement its model path and tool parser; verify the exact runtime and version with a harmless tool before connecting a browser.
Can I use standard Gemma 4 serving flags?
Do not assume so. DiffusionGemma needs diffusion-aware serving support and may require model-specific flags, templates, and parser configuration. Follow the instructions for the exact runtime build you install.
Does an open checkpoint mean a hosted API offers it?
No. Open weights and hosted provider availability are separate. Check the provider’s current model catalog, access terms, region, and tool support before selecting it.
Can the model reuse my website login?
The model does not hold the login. A local Tabbit CLI task can operate in the Tabbit environment where you have already signed in, if the launcher and local agent are connected to that instance.
What should I test first?
Use a harmless declared function and confirm the runtime parses and executes it once. Then use a read-only page task and check the saved report before allowing any write action.