Yes, offline AI models on your phone can chat in many languages besides English, but quality drops faster than it does with cloud chatbots. Major languages such as Spanish, French, German, Portuguese, Chinese and Japanese work well on a 4B model or larger. Less widely written languages work, but expect more grammar slips and odd word choices, especially on the smallest models.
Here’s what the current phone-sized models support, why smaller models struggle with some languages, and how to get better replies in yours.
Which offline models support which languages? #
| Model | Size on phone | Language coverage (per the model makers) |
|---|---|---|
| Qwen 3.5 (0.8B, 4B, 9B) | 0.8 to 5.7 GB | 201 languages and dialects |
| Gemma 4 E2B and E4B | 2.0 to 3.0 GB | Trained on 140+ languages; 35+ supported out of the box |
| Ministral 3 3B | 2.2 GB | Dozens, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic |
| GLM 4.6V Flash | 6.2 GB | Built by Zhipu AI with a focus on Chinese and English |
Sources: the model cards for Qwen 3.5, Gemma 4 and Ministral 3.
“Supports” is a wide word. A model trained on 200 languages has seen far more English, Chinese and Spanish than Amharic or Lao, and it shows. The best way to judge is to try your language on your phone.
Why do small models struggle with some languages? #
Less training text #
Models learn a language from the text they’re trained on. The web has vastly more English than, say, Swahili or Sinhala. A large cloud model has enough capacity to learn the rarer languages reasonably well anyway. A 1B or 4B model has to prioritize, and the rarer languages get squeezed.
More tokens per word #
Models read text in chunks called tokens. English words are often one token each. Words in many other scripts, including Devanagari, Thai, Amharic and Georgian, can take several tokens each. That has two practical effects on a phone:
- Slower replies. If a sentence costs three times as many tokens, it takes roughly three times as long to generate.
- Less fits in memory. The model’s context window fills up faster, so long conversations lose their early turns sooner.
Mixed-language drift #
Small models sometimes drift into English mid-answer, especially on technical topics, or mix in words from a related language. A system prompt that says what language to use fixes most of this.
Which model should you pick for your language? #
- Spanish, French, German, Portuguese, Italian: almost any model here works. Qwen 3.5 4B or Gemma 4 E4B give natural, fluent replies on an everyday phone.
- Chinese: Qwen 3.5 (built by Alibaba’s Qwen team) and GLM 4.6V Flash are strong. On an 8 GB phone, GLM is also good at reading Chinese text in photos.
- Japanese and Korean: Qwen 3.5 4B or 9B, or Gemma 4 E4B.
- Arabic, Hindi, Indonesian, Turkish, Vietnamese: start with Qwen 3.5 or Gemma 4, the two families with the widest coverage. Use the largest size your phone runs.
- Less-resourced languages: Qwen 3.5 9B if your phone has 6 GB of RAM or more. On smaller phones, keep expectations modest and prefer short tasks such as translating a sentence over long free-form chat.
Size matters more for non-English use than for English. The step from a 1B to a 4B model is dramatic for grammar and vocabulary in most languages.
How to get better replies in your language #
- Set the language in a system prompt. “Always answer in Brazilian Portuguese. Use informal tone.” Put it in the chat’s system prompt so it applies to every reply.
- Write your question in that language. Models usually answer in the language they’re asked in.
- Name the variety. “Mexican Spanish,” “European Portuguese,” “Simplified Chinese,” “Modern Standard Arabic.” Otherwise you get whichever variety dominated the training data.
- Lower the temperature for factual answers. A lower setting (around 0.3) makes wording more conservative and cuts down invented words.
- Ask for a check. “Is this sentence grammatical? Correct it and explain the changes.” Models are often better at correcting text than producing it from scratch.
- Keep conversations shorter in languages that use many tokens per word, and start a new chat when the topic changes.
What about translating on the go? #
Translating everyday phrases, messages, menus and signs is one of the most useful offline tasks, and a vision model can read text straight from a photo. We compare local AI with dedicated translation apps in offline translation on your phone.
Testing your language before you travel #
Before relying on an offline model abroad, spend five minutes on this test while you still have Wi-Fi:
- Download two candidate models.
- Ask each the same three things in your language: a casual question, a request to rewrite a short paragraph more formally, and a request to explain a simple concept.
- Compare fluency and speed. A live tokens-per-second readout makes the speed difference obvious.
- Keep the one that reads more naturally, and delete the other to free up storage.
Personal LLM makes this comparison straightforward. Its catalog includes Qwen 3.5, Gemma 4, Ministral 3 and GLM 4.6V Flash, each with a “Fits your device” badge based on your phone’s RAM, and every reply shows tokens per second. You can set a system prompt per chat (or a default for every new chat), and switch models mid-conversation from the chat header to hear how each handles the same question. Everything runs offline once downloaded, and you can have answers read aloud with on-device text-to-speech. For which model suits your hardware, see how much RAM you need to run an LLM on a phone.
Frequently asked questions #
Can offline AI speak my language? #
Probably. Current small models such as Qwen 3.5 cover around 200 languages and dialects, and Gemma 4 was trained on 140 or more. Quality varies a lot, though: major languages are fluent, and rarer ones are usable but less polished on small models.
Why does my local AI answer in English when I ask in another language? #
Small models lean toward English because most of their training text was English. Put “Always reply in [language]” in the system prompt and write your questions in that language, and the drift mostly stops.
Is offline AI slower in some languages? #
Yes. Many non-Latin scripts take more tokens per word, so the model generates more tokens for the same sentence and replies take longer. The model’s memory also fills sooner in long conversations.
What is the best offline AI model for Chinese? #
On phones with 8 GB of RAM or more, GLM 4.6V Flash and Qwen 3.5 9B are strong choices. On mid-range phones, Qwen 3.5 4B handles Chinese well at a smaller size.