Use Hy-MT2-30B-A3B for Browser Form Handoffs

Hy-MT2-30B-A3B can translate fields copied from a page, but it does not navigate a browser or submit a form by itself. Use it as the translation step, then give the reviewed values to a local agent that can access Tabbit CLI. The agent can read an authorized signed-in page, fill a test form, and stop before submission. Tencent’s exact weights ID is tencent/Hy-MT2-30B-A3B; its model card describes a translation model, not a tool-using browser agent.

In a Reddit thread about a manga translation tool, one user preferred HY-MT2 in Chinese and Korean tests but said its instructions needed to be followed. The tool maker said they had not tested it enough on manga to compare fairly and valued image context for fragmented dialogue. For form work, the same lesson applies: give the model the exact source text and target field labels, then inspect every mapping. These are individual comments, not a benchmark. Original discussion.

Tencent lists the 30B-A3B model as a translation model and documents Transformers, vLLM, and SGLang serving paths. Its card recommends explicit translation instructions and says there is no default system prompt. The card’s language badge says 36, its overview says 33, and its table lists 36 entries. The languages needed here, including English, Chinese, Spanish, Japanese, Korean, French, and Portuguese, appear in that table. Confirm the source-target pair with a non-sensitive sample. Hugging Face currently says the model is not deployed by an Inference Provider. The model card’s hy_v3 architecture label and MoE description do not establish tool calling. Tencent model card · Tencent source repository.

Key takeaways

• Hy-MT2-30B-A3B translates text; the local agent handles the browser.

• Keep field keys, source URLs, and values linked.

• Use synthetic data and a test page on the first run.

• A human checks every value before any submission.

Treat the model as a translator

Keep the data path narrow. Read only fields you are authorized to use, remove identifiers that the form does not need, and ask Hy-MT2-30B-A3B to translate the values while preserving their field keys. A local agent can then open the destination in Tabbit, fill a test draft, compare each populated value with its source, and stop before Submit. Do not paste personal data into an unapproved hosted model. If the selected agent client cannot pass model output into its local tools, hand the translated mapping to a separate local agent. Verify skill discovery, local launcher reachability, model serving, and website login independently.

Prompt for a field-safe translation.

Translate the values in this source-to-form mapping into Spanish. Keep every key unchanged and return valid JSON only. Do not infer missing values, modify IDs, or submit anything.

{"release_name":"Spring maintenance","window":"Thursday, 10:00 UTC","rollback":"Available from the release menu"}

The output format should preserve release_name, window, and rollback. Any translation shown in a demo should be marked illustrative unless it came from the selected model runtime. Use dummy values only.

Map values, not intent

A translation model cannot decide whether a value belongs in a field. The target page and labels define that mapping. Preserve each key and source reference so the reviewer can trace every field. If a value is ambiguous or missing, leave it blank and ask the person responsible; do not let the model invent a replacement.

Fill a test form and stop

Name the task Hy-MT2 source-to-form — synthetic release form. Ask the local agent to open an authorized source page in Tabbit, read only three approved fields from the synthetic test source, and put their translated values in the matching test form fields. It must report the source URL beside each value, show the unsent draft, and stop before Submit. A reviewer checks field alignment, dates, units, names, and missing values. Do not submit even if the form looks complete. This is a proposed workflow; the project’s signed-in fixture validated session reuse, but no Hy-MT2 inference or form-fill run was performed. The earlier genuine browser fixture returned HTTP 200 with heading Release notes review and record GEO-104 after a second local task reused the signed-in test session. It did not expose the three fields in this synthetic example. No Hy-MT2 inference or form submission was performed; the mappings below are suggested workflow steps.

The translation model is not the authority for whether a value belongs in a particular field. The local agent should preserve field keys and source links; a person should resolve ambiguity and approve any write. Keep the example on a test page with synthetic data. No result from the CLI login fixture proves that a third-party production form is safe, authorized, or compatible. Tencent’s LICENSE.txt lists Apache-2.0, but check the license file for the exact model revision you download.

Decide whether the handoff is safe

Use Hy-MT2-30B-A3B for translation inside a data-to-form handoff only when the language pair, model endpoint, and handling rules are verified. Let a separate local agent perform the browser read and fill, then require a human review before submission. If the workflow depends on the model itself calling browser tools, this setup is not established.

For the next setup check, see 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.

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.