To compare Laguna M.1 on a signed-in browser, first provision a working M.1 inference endpoint, then run the same read-only page task with M.1 and a Laguna XS baseline through a local agent connected to Tabbit CLI. M.1 weights are available under Apache 2.0, but Poolside's hosted Chat/API shutdown date has passed and the current OpenRouter catalog check showed no Poolside routes. This article gives a comparison method, not results from a model run.
Key takeaways
Laguna M.1 is a downloadable 225B-total / 23B-active model under Apache 2.0. Do not treat 23B active parameters as its memory requirement.
Poolside Chat says its service closed on October 31, 2026. Old OpenRouter model pages do not prove a live hosted endpoint.
Tabbit Browser holds the signed-in site session; a local agent uses Tabbit CLI to read pages. The model server and browser connection are separate.
A fair model check keeps the page, prompt, browser profile, permissions and output schema fixed. Compare evidence and action boundaries, not a benchmark score made from one task.
Confirm the model route before comparing
Poolside introduced M.1 in preview on April 28, 2026 and announced downloadable weights in the May 26 update to the same post. The current official model card describes a text model with 225B total parameters, 23B active per token, a 262,144-token context and reasoning around tool calls. It lists supported runtime paths and Apache 2.0. The current organization page also lists the base and post-trained BF16, FP8 and NVFP4 artifacts. (Poolside release post · M.1 model card · Poolside model organization)
Open weights and a hosted API are different things. Poolside Chat says it shut down on October 31 and points users to host the models themselves. OpenRouter still has a per-model page with old free-tier text, but the public model catalog returned no poolside/ IDs on November 10. Do not configure M.1 against a remembered endpoint. Provision a local or private inference service, test its exact model ID and tool-call parser, then connect your agent client. No M.1 endpoint was tested for this article. (Poolside Chat notice · OpenRouter M.1 page · OpenRouter current catalog)
| Layer | Current evidence | Decision before the test |
|---|---|---|
| Model weights | M.1 artifacts remain on Poolside's official Hugging Face organization; M.1 is Apache 2.0. | Select the exact base or post-trained variant and review its license and runtime instructions. |
| Model serving | Poolside's announced Chat/API shutdown date has passed; OpenRouter's current public catalog had no Poolside IDs. | Provide your own reachable inference endpoint. Do not assume an old free route works. |
| Memory and runtime | M.1 lists 225B total / 23B active parameters and BF16, FP8 and NVFP4 variants. | Size the host for the full selected checkpoint, runtime overhead, context and concurrent sessions. Active parameter count is not loaded model size. |
| Browser access | A separate Tabbit CLI fixture test verified reuse of a signed-in browser session. | Run the local agent where it can launch the CLI and select the intended Tabbit Browser profile. |
Keep the baseline and M.1 runs comparable
Choose a small test page with fixed, non-sensitive rows. Laguna XS 2.1 is a current XS-family baseline, but it is a separate model release and license family. Its official page lists OpenMDW-1.1, while M.1 is Apache 2.0. Do not merge their license terms or assume their runtimes have identical templates. (Poolside model organization)
Use one local agent client and the same tool permissions for both runs. Confirm that the client can select the exact model IDs through your configured endpoints. Before opening the signed-in dashboard, send one harmless completion to each endpoint and a harmless tool-call request that does not touch private data. If either result is malformed or the client cannot select a model, stop and fix serving before testing browser access.
For the browser runs, keep the page and expected values unchanged. Start one named task for the XS 2.1 baseline and another for M.1. Sequential runs are easier to compare and avoid competing for the same local inference resources. Record the output and tool receipt for each; a model switch changes the answer generator, not the Tabbit Browser profile or page state. The MCP versus CLI guide explains the local browser connection options. For one client-specific example, see the Codex setup guide.
| Control | Keep identical | Record for each run |
|---|---|---|
| Input page | Same read-only URL and fixed row IDs. | Page title, account/profile name, timestamp and row count. |
| Prompt | Same baseline, requested fields and no-write rule. | Exact prompt version and model ID. |
| Tools | Same Tabbit skill, CLI permissions and allowed browser actions. | Tool calls requested, returned, parsed or rejected. |
| Expected answer | Same status/date values and output columns. | Correct fields, missed rows, unsupported claims and source links. |
| Runtime | Record the specific variant and serving configuration. | Quantization, context limit, response time and any errors. |
Run the same page task with each model
The example below uses a synthetic page called Release review — GEO-104. It is a fixture, not a test performed with M.1 or XS 2.1. Use a page you are authorized to read, and do not enable edit controls for the test.
Use the currently selected model only: [Laguna XS baseline ID for run A, or Laguna M.1 ID for run B]. Do not change the model during this run.
In the current Tabbit Browser profile, open the authorized test page: [page URL]. Read only. Compare visible rows against this fixed baseline: [row IDs and expected status/date]. Return the same table columns for both runs: row ID, observed status, observed modified date, source URL, match/mismatch/unknown, and reason. Cite the exact page link for every row. Do not infer hidden data. Do not click Save, Resolve, Publish, Delete, Invite, or any control that changes data. Stop after returning a report in this run's separate task tab, named "Laguna comparison — [XS baseline or M.1]". This page and name are illustrative fixtures.Invoke Tabbit through the client's real skill entry. Some clients use /tabbit; Pi documents /skill:tabbit; others expose /skills or discover skills from a directory. Skill discovery, the local CLI launcher, the selected model endpoint and the website login are four distinct checks. See how to connect an agent to an existing browser and the Tabbit CLI setup guide. Do not claim that a cloud model runner has local browser access unless its local handoff is separately verified.
Use Tabbit Browser for the human review point
Tabbit Browser keeps the target site in the user's chosen profile. Tabbit CLI lets a local agent act on pages there, while separate task tabs keep the baseline and M.1 reports distinct. The model endpoint can change between runs; the browser session stays tied to the selected profile. A user can keep working in other tabs while the task runs, then reopen each source link to review the report.
The parent task recorded a local fixture test: a first named Tabbit CLI task received HTTP 401 before sign-in. After a short-lived test cookie was set, a second named task read the fixture heading “Release notes review,” ID GEO-104, and received HTTP 200 in the same browser environment. This test verifies session reuse only. It did not use M.1, XS 2.1, an external agent or a real dashboard. The comparison prompt above is an illustrative procedure, not a measured result. (Tabbit CLI guide · Browser profiles and task tabs · What an agentic browser does)
Compare what each run actually read
Open the sources yourself and check the output against visible values. A row is a match only if its stable ID, status and date agree. Count missing and incorrect rows separately. A valid tool call that opens the page is not evidence that the model understood the content, and a correct answer from one small fixture is not a general benchmark.
| Result field | XS baseline | M.1 | Human review |
|---|---|---|---|
| Rows expected / returned | Fill from the fixed fixture. | Fill from the same fixture. | Recount visible rows; note omissions and duplicates. |
| Status and date accuracy | Mark correct, incorrect or unknown per field. | Mark the same way. | Reopen the source link and compare literal values. |
| Source link fidelity | Record whether each URL opens the cited row. | Record the same. | Remove any claim without a matching source. |
| Tool boundary | Log attempted and completed actions. | Log attempted and completed actions. | Confirm both runs stayed read-only. |
| Runtime cost | Record measured time and resource use if available. | Record measured time and resource use if available. | Keep measurements local to this setup; do not generalize. |
Community posts provide useful test questions, not a verdict. One LocalLLaMA commenter using M.1 said it “followed directions better” and that its agent flow worked better, while noting slow inference and a limited context window. An X user who tried several free models called M.1 their most compelling option for coding-agent and software-engineering work. A separate YouTube reviewer titled a short review “Promising Coding Agent, Mixed Benchmark” and described a local benchmark plus self-hosting effort. These are individual coding impressions. None measures signed-in browser accuracy, and their hosted routes are historical. (Reddit discussion · X post · YouTube review)
If the two runs disagree, first rule out different endpoint variants, runtime templates, account sessions and context limits. If a tool call fails, inspect the model parser and client support before blaming Tabbit. The browser first-task checklist covers the page and permission checks; the source-linked browser research guide shows how to review citations.
Use this comparison only after the M.1 endpoint and tool calls work in your own environment. Keep both reports separate, inspect every source link, and leave any later dashboard update to a person with authority to make it.
FAQ
Can Laguna M.1 read a signed-in browser without another tool?
No. M.1 is a model, not a browser session. A local agent client must call a running M.1 endpoint, invoke Tabbit CLI, and use the Tabbit Browser profile where the site is already signed in.
Are Laguna M.1 weights available to download?
Yes. Poolside's current model card and organization page list M.1 weights under Apache 2.0. Check the exact variant and repository license before using or redistributing it.
Can I still use Poolside's hosted M.1 API?
Poolside Chat says it shut down on October 31, 2026. The OpenRouter model page remains visible, but the public catalog checked on November 10 had no Poolside model IDs. Confirm a working route before configuring a client.
Does M.1 need only 23B parameters of memory?
No. The card lists 225B total and 23B active parameters. Active parameters per token do not mean the full checkpoint can fit in memory as a 23B model. Check the selected quantization, runtime, memory and offload plan.
How do I make a fair M.1 versus XS comparison?
Use the same page, prompt, baseline values, browser profile, tool permissions and output fields. Run each model separately in a named task, then compare source links and visible values rather than relying on an impression or one call.
Does the example modify the dashboard?
No. The task is read-only and uses a clearly labelled fixture. The agent should stop after returning a report, and a person should verify it in the browser before authorizing any later edit.