To use Muse Spark Open with a signed-in browser, first choose how to serve the model, then connect a local agent to Tabbit CLI. Meta distinguishes its hosted Muse Spark service from Muse Spark Open, a separate downloadable-weights release (Meta's developer documentation). For a team brief, ask the agent to read approved project pages in Tabbit Browser, attach a source URL to each fact, and stop at a document draft. A Reddit developer described getting the 816B checkpoint running on ten A40 GPUs, which shows why downloading weights and serving them are separate steps (deployment report).
Key takeaways
Hosted Muse Spark and the open-weight Muse Spark Open are different routes. Check what endpoint your client calls.
Open weights do not mean a small download, easy deployment, or unrestricted license.
Model hosting, the local agent, and Tabbit Browser each need their own access check.
Gather only approved pages, cite each detail, and draft without sharing.
A Reddit user cautioned that aggregate model scores do not predict your own task. Test the exact workflow and review its sources (discussion).
Choose the inference route before connecting the browser
| Route | What it means | Check before use | Main trade-off |
|---|---|---|---|
| Meta hosted Muse Spark | Send requests to Meta's hosted model API | Account, region, data handling, model ID | Simple to try; external service processes the request |
| Muse Spark Open weights | Download and serve the separate checkpoint | License, supported runtime, memory, endpoint | Control over serving; setup and hardware burden |
| Local browser agent | Runs Tabbit CLI beside Tabbit Browser | Skill discovery and local process permission | Reuses the local signed-in environment; cannot be assumed from a remote runner |
Meta describes Muse Spark Open as an 816B-total, 42B-active multimodal checkpoint and provides download instructions. The published Muse Model License includes attribution and terms for model-as-a-service. Review the license for your actual deployment instead of treating “open weights” as “open source.” The research environment's pi --offline --list-models muse-spark-open found no matching entry, and no inference was attempted. The installed Tabbit skill uses Pi's /skill:tabbit; that does not mean Pi hosts the model. See the Tabbit CLI quick start and complete guide.
Draft a team brief from approved pages
In Tabbit Browser, sign in to the team workspace yourself and open only pages the model endpoint is allowed to process. Have the local agent collect a decision, date, and source link from each page. Leave missing or conflicting details unresolved. One community thread about model indexes makes the same useful distinction for evaluation: an aggregate number does not tell you how the model will handle your own material.
Read the currently open, approved project pages in Tabbit Browser. Do not open private pages outside this task, change content, or share anything.
Prepare a document draft named "Muse Spark Open Team Brief — GEO-418". For each point, include the displayed decision, date, and source URL. Separate facts from interpretation; list conflicting dates or missing owners as unresolved.
Do not publish or invite recipients. Stop and show me the draft, citations, and unresolved items for review.GEO-418 is an illustrative draft name, not a real team record or a model run. Before using real company material, check the chosen host's retention and access controls. Local model weights do not automatically mean every connected tool or browser action stays local.
Review the source trail before sharing
| Draft claim | Evidence to keep | Human review |
|---|---|---|
| Decision | Exact visible text and page URL | Confirm it is current and approved |
| Date | Page's displayed date | Resolve disagreements; do not guess |
| Owner | Only if visibly listed | Confirm the person and sharing scope |
| Recommendation | Source facts plus the agent's summary | Separate inference from the original decision |
For setup, read how to connect an agent to an existing browser, the first-task checklist, MCP vs CLI, and Codex with Tabbit CLI. If your local client cannot run the skill, use the browser agent CLI tutorial. Tabbit Browser is a practical choice when you want a local agent to work in a browser session you already signed into. Related background covers browsers for AI agents, Playwright browser automation, and browser automation agents. Parent-run fixture evidence showed local session reuse only, not Muse model integration. Review the security risks of browser agents, verify every cited page, and let a person choose recipients and publish.
FAQ
Is hosted Muse Spark the same as Muse Spark Open?
No. Meta lists hosted Muse Spark on its Model API and publishes Muse Spark Open as a separate checkpoint with downloadable weights. Verify which route your client actually uses.
Does open-weight mean Muse Spark Open has an unrestricted license?
No. Meta publishes the Muse Model License. Read its attribution and model-as-a-service terms before deployment, especially if you redistribute or host the model for others.
Can I run Muse Spark Open on a regular laptop?
Do not assume so. The checkpoint is very large, and serving it depends on memory, hardware, quantization, and supported inference software. A hosted provider may be more practical.
Can Muse Spark Open use my existing Tabbit login by itself?
No. A local agent must run Tabbit CLI in the browser environment where you signed in. Model weights and browser access are separate.
Should the agent publish the team document?
No. Have it prepare a named draft with source URLs and unresolved items. A person should check privacy, recipients, citations, and sharing permissions before publishing.