Hy-MT2-7B can translate a short release-note draft, but it does not operate a browser by itself. Give the text to the model, then pass its suggested wording to a local agent that can call Tabbit CLI. The agent can open the signed-in source, compare the draft, and stop at a named review checkpoint. Tencent publishes the exact weight ID as tencent/Hy-MT2-7B; it is a translation specialist, not a browser agent.
A Reddit commenter who tried HY-MT2 said it worked better than Gemma 4 E2B in their Chinese and Korean tests, while also warning that the instructions mattered. Another participant said they had not tested it enough on manga to judge fairly and valued image context from a multimodal model. Those are useful reminders to test the same text and keep context visible, not benchmarks. Read the original discussion.
Tencent’s model card identifies the 7B model as a multilingual translation model and documents Transformers, vLLM, and SGLang serving paths. It recommends a translation prompt and says the model has no default system prompt. The Hugging Face page lists 36 languages, while its overview says 33 and the table lists 36 entries. English, Simplified Chinese, Traditional Chinese, Spanish, Japanese, Korean, French, and Portuguese are all in that table. Verify your exact language pair on a small, non-sensitive sample before processing a long document. The model card currently says the model is not deployed by a Hugging Face Inference Provider. Tencent model card · Tencent source repository.
Key takeaways
It is a translation specialist, not a browser agent.
Use the exact weights ID and verify the runtime and language pair.
Keep the source and proposed wording visible side by side.
Stop at a named review checkpoint; a person approves edits.
Who does what
| Part | Job | Boundary |
|---|---|---|
| Hy-MT2-7B | Suggest translated wording. | Does not browse or approve the text. |
| Local agent | Open the source and show both versions. | Needs verified Tabbit CLI access. |
| Tabbit Browser | Keep the signed-in site available locally. | Sign-in does not prove model access. |
| Reviewer | Check and approve the draft. | Owns the final wording decision. |
Use Hy-MT2-7B for wording, not browser control
Start with a short release note whose source and target are already approved. Ask Hy-MT2-7B for a translation only. Do not ask it to click, browse, or choose a site. If your local agent client can select the model and expose its text result to a local tool-using agent, verify that path with harmless text first. Otherwise, copy the proposed translation into the local agent yourself. Client skill discovery, launcher access, and website sign-in are separate checks. In clients with named skill commands, use the syntax that client documents; Pi uses /skill:tabbit, and some clients use /skills to discover available skills. Then install Tabbit Browser, sign in to the intended profile, and confirm the local agent can reach Tabbit CLI with diagnose.
Prompt for a controlled draft.
Translate the following release note into Simplified Chinese. Return only the translation. Preserve the date, time, product name, and meaning of the rollback instruction. Do not add claims or explanations.
Release window: Thursday, 10:00 UTC. Rollback is available from the release menu.Illustrative only, not a Hy-MT2 output: “发布时间：周四 10:00 UTC。可在发布菜单中回滚。” Run the prompt with your chosen local serving path and inspect its actual result before using it.
Translation review checklist
| Check | Accept when | Stop and fix when |
|---|---|---|
| Meaning | Instruction and negation match the source. | A rollback condition changes. |
| Details | Names, date, time, and units match. | A value is missing or reformatted ambiguously. |
| Wording | Reads naturally for the target audience. | It adds a claim that is not in the source. |
Set up the local browser path
The model, agent client, local launcher, and signed-in site are separate components. Check each one. Discover the installed tabbit skill using the client documented entry point, confirm the launcher works, and sign in to the target profile in Tabbit Browser yourself. Some clients use /tabbit; Pi uses /skill:tabbit, while clients that support it use /skills to discover skills. A remote model or runner does not gain access to the local browser just because it can produce a task plan.
Review a paired translation draft
Create a review task named Hy-MT2 translated draft — release note 104. Use a separate synthetic release-note source, because the genuine GEO-104 login fixture contains only its heading and record. Ask the local agent to open that approved source in Tabbit, show the source and model-suggested Chinese side by side, and save nothing. The reviewer checks the date, time, product names, negation, and rollback meaning, then approves or edits the wording. Stop at that visible checkpoint. A task tab keeps the check separate while you continue using your regular browser tabs. The genuine browser fixture result was HTTP 200 with the heading Release notes review and record GEO-104 in a second local Tabbit CLI task after the test sign-in. That was the only genuine result here. The fixture did not contain the release-note sentence used above; that sentence and its illustrative translation are synthetic, not a Hy-MT2 output.
Do not treat fluent output as verified translation. The Reddit thread included a commenter cautioning that a translation can look correct while being subtly wrong. For product or legal language, a bilingual reviewer should compare the translation with the source. Do not send unreleased or customer text to a hosted endpoint unless your organization has approved that route. The weights are Apache-2.0 according to the model’s LICENSE.txt, but the license does not establish a provider integration or permit disclosure of confidential text.
Decide whether this model fits
Use this workflow when you need a first-pass translation of stable, short copy and want a person to review it against the signed-in source page. If you need the model to navigate, fill fields, or decide what to click, choose a separately verified tool-using model and agent host. Hy-MT2-7B can supply wording; Tabbit CLI supplies local browser access.
Connect the next step to the right guide: Tabbit CLI guide, connect an agent to an existing browser, the first browser-agent checklist, browser MCP vs CLI, Codex and Tabbit CLI setup, browser-agent CLI tutorial, how browser agents control a browser, reuse a signed-in browser session.
FAQ
Can Hy-MT2-7B control Tabbit Browser?
No native browser or tool-calling connection is established by Tencent’s model card. Use it for translation, then pass its text to a separately verified local agent or a human.
What exact model ID should I select?
For the official Hugging Face weights, use `tencent/Hy-MT2-7B`. A provider may expose another alias, so verify the selected endpoint and model ID in that provider’s current documentation.
Which languages can I test?
The model card table includes English, Chinese, Traditional Chinese, Spanish, Japanese, Korean, French, and Portuguese. The page’s badge and overview disagree on the total count, so check the current language table and test your direction.
Does sign-in happen through Hy-MT2-7B?
No. Sign in to the target website in the Tabbit profile yourself. The local agent can reuse the browser session only if its Tabbit CLI connection is verified.
How do I review the translation?
Compare the exact source and output side by side. Check names, dates, numbers, negation, and instructions before the reviewer approves or edits the named draft.
Can I use a hosted inference endpoint?
The model can be self-hosted through documented runtimes. The Hugging Face page currently says it is not deployed by an Inference Provider. Check any other service’s model ID, data handling, and availability before sending text.