North Small Translate can translate text taken from a signed-in website, but it is not a browser agent. Use Tabbit Browser for the page session and a local agent with the Tabbit skill to read approved content. Send only those excerpts to North Small Translate, then review the translated brief before saving it. Cohere documents the model as a machine-translation model, with the API ID north-small-translate-1.0; its open weights are licensed for non-commercial use. The model card does not make the model itself a browser controller.
A commenter on the North Small Translate Reddit launch thread asked how it performs “outside benchmark-style translation” and called out the hardware trade-off behind 218B total and 25B active parameters. That is a useful reminder for this workflow: a multilingual model still needs a task-specific check. A separate Ollama translation discussion describes different results across Russian-English models and sizes. Those reports are personal experiences, not a North Small Translate benchmark.
Key takeaways
North Small Translate translates text. It does not automatically navigate, click, or authenticate to a website.
Keep browser reading in a local agent that can run Tabbit CLI on the same machine as Tabbit Browser.
Pass the model only text you have permission to share, with stable IDs and source URLs kept outside the translation output.
Verify language pair, names, dates, negation, numbers, and links against the original page.
Cohere’s hosted API is the practical path for many users. Open weights do not imply lightweight local inference.
North Small Translate and Tabbit at a glance
| Job | Use | Boundary to check |
|---|---|---|
| Read a logged-in page | Tabbit Browser through a local CLI-capable agent | The agent must use the intended Tabbit profile and site session |
| Translate approved excerpts | north-small-translate-1.0 through Cohere Chat V2 or an eligible local deployment | The translation model sees only text you pass to it |
| Keep source provenance | Local agent or your review sheet | Preserve the original URL, page title, and stable excerpt ID |
| Draft the brief | The local agent in a separate Tabbit task | Review a named draft before saving or sharing it |
First check that the model fits the text job
Cohere describes North Small Translate as an MoE model for machine translation across more than 50 languages. Its documented context is 16K tokens, and the model ID for Chat V2 is north-small-translate-1.0. The Hugging Face weights use CC BY-NC 4.0 and Cohere’s acceptable-use policy. Cohere lists suggested W4A16 hardware at two H100s or one B200, so “open weights” is not the same as “easy to run on a laptop.” Check the Cohere model guide and model card before choosing a hosted or local route.
Do not infer tool use from a general model capability menu. The model directory used for this assignment does not mark North Small Translate as natively supporting tool calls. If your client cannot select it as its text model, or if it cannot call local tools, split the work: one local agent reads the pages and exports approved text; North Small Translate translates the export; you or the agent then assembles the linked brief. Never assume the model can invoke Tabbit CLI just because the model is hosted in an agent UI.
Prepare the browser and separate the three checks
Install Tabbit Browser, sign in to the target site in the Tabbit profile you want to use, then start a local agent that can invoke Tabbit CLI. Tabbit’s CLI quick start documents /tabbit for clients that discover the installed skill. Check that the client actually lists the tabbit skill and can read its SKILL.md; some clients expose a Skills menu or /skills, while Pi documents /skill:tabbit. A command being visible proves skill discovery, not that the CLI launcher can reach the browser. Run tabbit-cli diagnose from the same machine to check that connection. Then open the target site in Tabbit and confirm its signed-in state yourself.
Keep those checks separate:
| Check | Evidence | What it does not prove |
|---|---|---|
| Skill discovery | The agent lists tabbit or accepts the client’s documented skill command | The browser controller is reachable |
| Launcher reachability | tabbit-cli diagnose reports a running controller | The target website is signed in |
| Site session | You verify the intended page in the correct Tabbit profile | The translation endpoint is approved for that content |
A local Tabbit CLI check in this environment verified session reuse with a fixture page: one task received an authorization error before sign-in, and another named task read the signed-in fixture. This confirms the Tabbit CLI session flow tested here. It does not test Cohere model hosting, the client’s model integration, or access to any third-party account.
Read the page with the local agent, then translate selected text
Start with public pages when you are collecting context. For a signed-in page, ask the agent to read only the specific fields needed for the brief. Keep sensitive fields out of its output. The agent should return a short list with a stable ID, title, URL, and excerpt. Before sending anything to North Small Translate, inspect that list and remove text that should not leave the browser environment.
Then ask the model to translate, not summarize, rank, or act on the material. A translation prompt should preserve source IDs and links while making the expected shape explicit:
Translate the approved excerpts into Simplified Chinese.
Translate only the text inside each excerpt. Do not browse, summarize, add facts, or change IDs and URLs. Preserve numbers, product names, dates, quoted terms, and negation. If a sentence is ambiguous, keep a short note in review_flags instead of guessing.
Input:
- id: page-01-paragraph-02
url: https://example.com/release-notes
text: The dashboard is read-only until an owner approves the change.
Return one JSON object per input item with id, url, translated_text, and review_flags.Expected fixture output:
{"id":"page-01-paragraph-02","url":"https://example.com/release-notes","translated_text":"在负责人批准更改前，控制台为只读状态。","review_flags":[]}The URL and paragraph above are test parameters. They illustrate format only. The local agent should keep the source list, translate only selected text, and compare the output with the source before drafting.
Build a linked brief without losing provenance
A brief should make it possible to reopen the source and check every claim. Keep one record per source excerpt. If two pages disagree, show both rather than asking the translation model to decide which one is current. A reviewer can then follow the link and resolve the conflict in the browser.
This is where a separate Tabbit task helps. The CLI creates task-owned browser tabs while you keep other tabs available for reading and review. You can ask the local agent to prepare a draft in a named file or draft area such as North Small Translate source brief. The agent should return the draft location and stop at a visible review checkpoint. Do not have it post, submit, or update a live record unless you explicitly approve that action.
If you need help choosing an access method, compare browser MCP and CLI. For permission boundaries, use the first-task checklist. The Tabbit CLI complete guide, agent connection guide, Codex setup, Tabbit browser overview, best practices, and agentic browser explainer cover the surrounding setup.
Review the translation as a separate step
A fluent translation can still reverse meaning or obscure a constraint. Review headings, dates, prices, names, statuses, and words such as “not,” “only,” and “until.” Open the original page in the browser, compare each translated excerpt by ID, and verify every URL still points to the source. If the page content is technical, ask a domain reviewer to check terminology before using the brief in a decision.
The Reddit discussion about local translation models illustrates why language pair and model size matter: users reported different quality across Russian-English tests and several model sizes. It does not establish how North Small Translate performs. Use a small test set from your own pages and keep the source, translation, and correction together. For sensitive content, route the text to an approved local deployment or do not send it to an external API.
The practical choice is straightforward. Use Tabbit CLI for the signed-in browser read, North Small Translate only for authorized translation, and a person for final verification. Download Tabbit Browser if you need a browser context that a local agent can access through the CLI.
FAQ
Can North Small Translate click or read browser pages?
No. Treat it as a translation model for text you provide. A separate local agent or you must read the page through Tabbit CLI, and a human should decide which content may be sent for translation.
What is the North Small Translate API model ID?
Cohere documents the Chat V2 model ID as north-small-translate-1.0. A model name in an API does not give it browser access or tool permissions.
Can I run North Small Translate locally?
Cohere publishes open weights under CC BY-NC 4.0, subject to its acceptable-use policy. The suggested hardware is substantial, so check the current model card and license before planning a local deployment.
Should I send a private page to Cohere for translation?
Only if your organization permits that data path and you have selected the correct deployment. The Cohere API is a hosted service; self-hosting the open weights changes the data path but still requires suitable hardware and license review.
How do I preserve links while translating a page?
Keep each source URL and text block under a stable ID, and ask the model to translate only the supplied text. Compare every output block with its source before writing a summary or draft.
Can a separate Tabbit task keep my browser available?
Yes, Tabbit CLI tasks own their tabs, so you can keep working in other windows while a task runs. Confirm that your local client can reach the Tabbit launcher and that the intended site is signed in first.