Use Macaron V1 Tall for a controlled browser task by routing a local agent through Mind Lab's Mixture-of-LoRA-Harness, then using Tabbit CLI to inspect two authorized signed-in pages. Compare a stable record ID and write the result to a separate draft; do not change either source page. Tall is the smaller V1 route intended for local, lower-latency, or on-premises use. Mind Lab describes Tall as a 50B release built on Qwen3.6-35B-A3B with four LoRA specialists; Venti is a separate 748B GLM-5.2-based release. (Mind Lab V1 README · release overview)
The official repository says full V1 serving uses its Mixture-of-LoRA-Harness in front of vLLM or SGLang. Confirm the current hardware, dependencies, license terms, and tool-call behavior before using real pages. Tall's model card lists MIT metadata, while Mind Lab notes that the model also inherits the Qwen base model's terms. (Tall model card)
Key takeaways
Tall is not a Venti alias. Pick it for the smaller local/on-prem route and use Venti when its distinct hosted flagship route fits.
The local checkpoint ID is
mindlab-research/Macaron-V1-Tall; the current hosted API ID ismacaron-v1-tall. Verify which endpoint your client is using.For the complete V1 setup, follow the current MoL harness instructions. A plain base-model server does not imply the four-specialist workflow is active.
Tabbit Browser holds the signed-in website session. A local agent calls the Tabbit skill separately from the model endpoint.
Compare by stable ID, cite both source pages, and leave a separate draft for a person to review.
Choose the serving route
| Route | Model identifier | Fit | Check before use |
|---|---|---|---|
| Local V1 harness | mindlab-research/Macaron-V1-Tall | Local, on-premises, or latency-sensitive work | Current MoL setup, compute, dependencies, base-model terms, and tool calls |
| Mind Lab hosted API | macaron-v1-tall | Managed endpoint where policy permits hosted processing | Account access, current model ID, privacy boundary, and client tool support |
| Venti route | macaron-v1-venti | Separate flagship option with a larger GLM-5.2 base | Do not assume Tall's latency, cost, or deployment profile applies |
Mind Lab lists Tall and Venti as different models and currently documents both a hosted API and a local harness. The API ID is not the Hugging Face repository name. If records contain private information, use the local route only after confirming your local deployment and data handling; a hosted API sends request content to that service. See the official model list and local harness repository.
Prepare a read-only comparison
Sign in to both source sites yourself in Tabbit Browser, then ask your local agent to open only those authorized pages through Tabbit CLI. Keep the lookup key stable and use a synthetic example such as TALL-REC-0042 while validating the workflow. Record exact titles, URLs, visible timestamps, and relevant field values. Do not paste session cookies or credentials into the prompt.
| Result | Evidence to capture | Next step |
|---|---|---|
| Match | Same ID and same compared values in both pages | Cite both pages and mark the checked fields |
| Conflict | Same ID but one or more values differ | Quote both values and ask an owner to resolve |
| Missing on one side | ID appears in only one authorized page | Report which page was checked; do not infer deletion |
| Not verified | Page inaccessible or field not visible | Mark unknown and request human follow-up |
The record state is a comparison result, not permission to edit. Use a separate named draft and make every statement traceable to its source. The browser research with sources guide explains how to keep links attached to claims; the profiles and task tabs guide covers separate browser contexts.
Configure the local agent and browser skill
Use the harness's current OpenAI-compatible endpoint instructions with a client that supports the required tool calls. OpenCode's provider configuration supports custom OpenAI-compatible endpoints, and its skill system invokes an installed skill by mentioning @skill-id. Check that the skill ID is actually tabbit before using @tabbit. (OpenCode providers · skills)
Tabbit CLI provides the local browser connection. Its quick start and existing-browser guide describe session reuse. The first-task checklist covers safe initial tool checks, while MCP vs CLI helps select a connector. The agentic browser overview explains why the browser session remains separate from the model.
Before opening private pages, run a harmless test against a public page: ask the agent to call the browser tool and return the page title. Confirm that the active endpoint is Tall and the harness has loaded the intended experts. If the tool call fails or the selected route is unclear, stop before accessing the workspace.
Prompt for a stable-ID check
The following prompt uses placeholders and a synthetic ID. It is a proposed workflow, not a completed Tall integration test.
In OpenCode, use @tabbit and start a separate task named TALL-REC-CHECK-01.
Use the configured local Macaron V1 Tall endpoint through the approved V1 harness.
Open only these two pages that I am authorized to access:
[SIGNED-IN PAGE A URL]
[SIGNED-IN PAGE B URL]
Search both pages for the exact stable ID TALL-REC-0042.
For each page, record the title, URL, visible timestamp, and values of:
[FIELD 1], [FIELD 2], [FIELD 3]
Classify each field as match, conflict, missing, or not verified.
Do not edit, submit, share, or delete either source record.
Create a separate document named "Record reconciliation, draft 01".
Link each finding to the source page and quote only the needed field values.
Show the draft with both links and stop for human review.Adapt the source scope and expected fields using this browser-agent prompt template.
Review both pages yourself before sharing the result. A missing value can mean it was not visible or the page could not be read; do not treat it as proof that the record was deleted. If the pages disagree, preserve both values and ask the responsible owner.
Evidence and limitations
One Macaron V1 discussion includes a reader saying they lacked the compute to try the model, and another participant points to the required LoRA harness. These are individual community observations, not proof of model quality or compatibility. The original discussion also contains skepticism about reported benchmark scores. I found no relevant X post or YouTube comment that I could verify, so this recommendation rests on official model/setup documentation and a read-only workflow design.
The workflow is appropriate when a team can operate the documented local harness, authorize access to both pages, and review the draft. It is not a promise that Tall will resolve ambiguous records correctly. Check fields against both source pages before acting.
FAQ
Is Macaron V1 Tall just another name for Venti?
No. Mind Lab describes Tall as a 50B model based on Qwen3.6-35B-A3B plus four LoRA specialists, while Venti is a 748B GLM-5.2-based model. They have different model IDs and serving profiles.
Can I serve the full Tall model as an ordinary base checkpoint?
Mind Lab's V1 README describes the full V1 workflow using Mixture-of-LoRA-Harness in front of a serving engine. Follow the current harness instructions and verify the chosen route; a base-model example is not automatically the complete V1 setup.
What is the local model ID?
The Hugging Face repository is `mindlab-research/Macaron-V1-Tall`. The hosted API catalog uses `macaron-v1-tall`; these identify different routes.
Does Tall connect to my signed-in pages?
No. A local agent must separately use Tabbit CLI to reach the authorized session in Tabbit Browser. The model does not receive browser cookies by itself.
Can the agent update the two records after comparing them?
This workflow is read-only. Ask it to report matching, missing, or conflicting fields in a separate document and stop for human review.
Does this prove Tall works end to end with Tabbit?
No. Official docs establish the model and harness routes plus Tabbit's local browser connection separately. This specific Tall-and-Tabbit workflow has not been tested end to end.