Quick start
- Open Tools in Noema.
- Under Language, open Translation.
- If prompted, download a compatible Hy-MT2 model. Noema only shows models that fit the device's current memory budget.
- Choose a Source language, or leave it on Detect language.
- Choose the Target language.
- Enter text in Original, or use the clipboard button to paste it.
- Leave Preserve formatting on when line breaks, markup, code, or placeholders matter.
- Select Translate. With a hardware keyboard, you can also press Command–Return.
- When the translation is complete, select Copy to place it on the clipboard.
You can cancel a translation while it is running. For multi-paragraph text, Noema shows paragraph progress in Simple and Advanced modes.
First-time setup
Translation uses a dedicated local model. The first time you open the tool, Noema evaluates the device's current app memory budget and offers only compatible options.
In Beginner mode, setup is simplified:
- Noema chooses the smallest Hy-MT2 model that passes the memory check.
- The setup page shows the one-time download size before downloading.
- The download can be paused, resumed, cancelled, or retried.
- Noema checks available storage and waits for a usable connection when necessary.
In Simple and Advanced modes, you can choose among every model that currently fits. Noema marks the largest fitting option as Recommended for quality, while keeping smaller models available for faster translation and smaller downloads.
Noema recalculates the list from the device's reported app memory budget and checks the fit again immediately before starting a download.
Translation models
Noema supports three official Hy-MT2 GGUF options. All three use the Q4_K_M quantization supported by Noema's local llama.cpp runtime.
| Model | Approximate download | Noema's guidance | Important detail |
|---|---|---|---|
| Hy-MT2 1.8B | 1.13 GB | Fastest; suited to everyday translation | The automatic choice in Beginner mode when it fits |
| Hy-MT2 7B | 4.62 GB | Higher quality; better nuance and instruction following | Requires substantially more storage and memory than 1.8B |
| Hy-MT2 30B-A3B | 18.24 GB | Maximum quality for complex translation | Currently runs CPU-only in Noema and can be considerably slower |
Download size is not the same as runtime memory use. Noema's fit check includes model weights, an 8,192-token context, the KV cache, runtime buffers, and conservative working headroom. A model can therefore be hidden even when the device has enough free storage for its file.
There is no universal “best” model for every device. Start with 1.8B when responsiveness matters, use 7B when the device can fit it and nuance matters more, and reserve 30B-A3B for memory-rich devices and work where its much slower CPU-only execution is acceptable.
Supported languages
Any supported language can be selected as the source or target. The source can also be detected automatically.
- Arabic, Bengali, Burmese, Cantonese
- Chinese, Traditional Chinese, Czech, Dutch
- English, Filipino, French, German
- Gujarati, Hebrew, Hindi, Indonesian
- Italian, Japanese, Kazakh, Khmer
- Korean, Malay, Marathi, Mongolian
- Persian, Polish, Portuguese, Russian
- Spanish, Tamil, Telugu, Thai
- Tibetan, Turkish, Ukrainian, Urdu
- Uyghur, Vietnamese
Language names follow the language selected for Noema's interface when a localized system name is available. Automatic detection works best when the source is primarily one language. For short, ambiguous, or mixed-language text, choosing the source language explicitly can improve consistency.
Noema disables translation when the explicitly selected source and target are the same.
Preserving formatting
Turn on Preserve formatting when the structure of the source matters. The model is instructed to preserve:
- Line breaks and paragraph boundaries
- Delimiters and placeholders
- Markup
- Code
- Other non-user-facing structure
Noema separates the source at blank-line paragraph boundaries, translates each paragraph independently, and then restores the original blank-line separators exactly. A translation is shown only after every paragraph has returned a usable result, which prevents a partially translated document from being presented as complete.
Formatting within a translated paragraph is still model-generated. Always verify code, markup, variables, URLs, product names, and placeholders before using the result in production.
Turn Preserve formatting off when you want the model to use more natural formatting in the target language.
Translating longer text
Noema processes multi-paragraph text one paragraph at a time. This has several benefits:
- Long documents do not have to fit into the model as one request.
- Existing blank-line structure can be retained.
- Progress can be reported paragraph by paragraph.
- A malformed paragraph can be retried without accepting an incomplete final document.
Each paragraph still has to fit inside the translation session's bounded context. If Noema reports that the text is too long for the current context, divide an unusually long unbroken paragraph into smaller paragraphs with blank lines, then translate again.
The Translation workspace currently accepts text entered or pasted into the editor. It does not directly import a PDF, Word document, image, or webpage. Extract or copy the text first, then review the translated formatting before placing it back into the original document.
Automatic detection and language swapping
Choose Detect language when you do not know the source language. Each translated paragraph reports a detected source language, and Noema uses the most frequently detected language for the completed result.
After a translation finishes, select the swap button to:
- Make the previous target language the new source language
- Use the detected or selected source language as the new target
- Move the translated output into the Original editor
This is useful for translating a result back in the opposite direction. A round-trip translation is a convenience check, not proof that either translation is exact.
Privacy and offline use
The Translation workspace only accepts a recognized local Hy-MT2 GGUF model. It does not route translation text through a configured remote endpoint, a cloud model, or the chat model's provider.
- During model setup: Noema connects to Hugging Face to download the selected Hy-MT2 model file.
- During translation: The text is processed by Noema's embedded local llama.cpp runtime on the device.
- After setup: Translation can run without an internet connection, including in Off-Grid mode.
- Chat history: Translation requests and results are not inserted into a chat conversation.
The workspace does not provide a saved translation history. Copy any result you want to keep.
How translation affects the chat model
Noema treats Translation as an isolated workspace. It may temporarily load the selected Hy-MT2 model into the local inference runtime, but it does not replace your normal chat-model preference.
- In Simple and Advanced modes, the translation model remains available while you stay in the workspace. Your previous chat model is restored when you leave.
- In Beginner mode, Noema returns to the managed primary model after a translation finishes or is cancelled.
- Entering Beginner mode while a translation is active cancels the translation and applies Beginner's managed runtime policy.
Loading and restoring large local models can take time. This is expected and is separate from the translation itself.
Tips for better translations
- Select the source language instead of automatic detection for very short text, names, or multilingual content.
- Keep Preserve formatting on for templates, localization strings, code-adjacent text, or content with placeholders.
- Turn it off for ordinary prose when natural target-language formatting matters more than source layout.
- Split extremely long paragraphs at natural boundaries.
- Use the 7B model when nuance matters and the device offers it safely.
- Use the 1.8B model when speed and lower memory use matter most.
- Treat the 30B-A3B model as a quality-focused, CPU-only option rather than a fast default.
- Check terminology, proper nouns, measurements, dates, numbers, and negation manually.
- Keep an untouched copy of the source when translating structured content.
Troubleshooting
| What you see | What it means | What to do |
|---|---|---|
| No Hy-MT2 model fits safely | Even the smallest supported model is above Noema's reported app memory budget. | Select Check Again to recalculate the fit. If no model appears, private translation is not available under that device's reported budget. |
| Private translation isn't available on this device | Beginner mode could not find a compatible translation model that safely fits. | Select Check Again. Use another supported device if the result does not change. |
| Free some storage, then try again | The one-time model download does not have enough working storage. | Free local storage and retry. The required free space can be larger than the final model file while a download is in progress. |
| Waiting for a connection | The model is not installed and the download cannot currently proceed. | Reconnect, then select Try Again. On a constrained or metered connection, confirm the download when Noema asks. |
| The Translate button is unavailable | The input is empty, no compatible model is installed, the target is missing, the explicit source equals the target, or a translation is already running. | Check the text, language pair, model setup, and current progress. |
| This text is too long for the translation model's current context | At least one paragraph cannot fit in the bounded translation context. | Break the long paragraph into smaller paragraphs with blank lines and retry. |
| Failed to load the translation model | The local runtime could not load the selected model and settings. | Free memory, retry, or choose a smaller model from Models. |
| The model returned an unreadable translation | The model did not return a complete structured paragraph, even after Noema retried it once. | Try again. If it repeats, shorten the paragraph or choose another fitting model. |
| The result changed markup or placeholders | Formatting preservation is model-assisted and is not a formal parser. | Turn on Preserve formatting, translate smaller sections, and compare the result with the source before use. |
| Translation feels slow | Model loading and local generation depend on the model and device. The 30B-A3B option is CPU-only. | Keep the workspace open between translations or choose a smaller model. |
Current limitations
- Translation is text-only inside this workspace.
- The tool does not directly import document files, images, audio, or webpages.
- Only the listed 38 languages are available in the language selectors.
- There is no built-in glossary, translation memory, or terminology lock.
- There is no persistent translation history.
- Formatting preservation cannot guarantee byte-for-byte preservation inside model-generated paragraphs.
- Automatic detection can be ambiguous for short or mixed-language input.
- The 8,192-token context applies per paragraph, not to an unlimited single block of text.
- Translation quality varies by language pair, model size, subject, and writing style.
Frequently asked questions
Does translation require an internet connection?
Only for the model download. Once a compatible Hy-MT2 model is installed, the translation itself runs locally and can work offline.
Is my text sent to Tencent or Hugging Face?
Noema contacts Hugging Face to download the model file. The Translation workspace then runs that model locally; it does not send the text to Tencent, Hugging Face, or a configured remote inference provider.
Why do I see fewer model choices than another device?
Noema filters the catalog using the current device's app memory budget and a conservative runtime estimate. A device with a larger safe budget can show larger models.
Why is the largest model not always available?
The 30B-A3B download is approximately 18.24 GB, requires much more runtime memory than its file size alone suggests, and currently runs CPU-only in Noema. It only appears when the device passes the memory-fit check.
Which model should I choose?
Choose 1.8B for speed and broad device compatibility, 7B for stronger nuance when it fits, and 30B-A3B only when maximum quality matters more than speed. Noema's Recommended badge identifies the largest option that currently passes its safety check.
Can Translation use my OpenAI, OpenRouter, Ollama, LM Studio, or Constellation model?
No. This workspace deliberately uses one of Noema's recognized local Hy-MT2 GGUF models.
Does Translation change my default chat model?
No. The Hy-MT2 runtime is temporary. Noema restores the applicable chat runtime when the Translation workspace ends, with additional managed restoration in Beginner mode.
Are translations saved in my conversations?
No. They are not inserted into chat history. Use Copy if you want to keep or move a result.
Can I translate a whole document?
You can paste multi-paragraph text, and Noema will process it paragraph by paragraph. Direct file import and layout-aware document export are not currently part of the Translation workspace.
Can I trust the output without review?
No machine translation should be assumed perfect. Review important output, especially content involving safety, health, law, money, identity, deadlines, or contractual obligations.
How it works
Noema's translation pipeline is designed to favor complete, inspectable results:
- The app verifies that the selected model is a supported Hy-MT2 GGUF.
- It divides the source into paragraphs at blank-line boundaries and assigns each paragraph a stable internal identifier.
- It sends each paragraph to the local model with the chosen language direction and formatting policy.
- It requires a structured result containing the matching paragraph identifier, the translation, and the detected source language.
- It rejects empty, mismatched, malformed, or token-limited output and retries the paragraph once.
- It assembles the completed paragraphs with the source's original blank-line separators.
- It publishes the final translation only if every paragraph is present and non-empty.
Source content is explicitly marked as content rather than instructions. The translation calls are also treated as auxiliary local work, so they do not create durable chat-generation checkpoints.