DeepSeek Languages: Natural and Programming Language Support

If you are deciding whether to ship a multilingual product on DeepSeek, the practical question is not “does it speak French?” but “how well, on which model, at what cost, and with what failure modes?” This guide covers what DeepSeek itself documents about the languages its models handle — natural languages and programming languages — on the current API models, deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813), plus the prompts and parameters that make the biggest difference.

The short answer: what DeepSeek actually supports

DeepSeek does not publish a closed list of “officially supported” natural languages the way some translation APIs do, and it publishes no per-language quality ratings. Its models are general-purpose multilingual models: the DeepSeek-V3 technical report (December 2024) says the pre-training corpus expanded “multilingual coverage beyond English and Chinese”, and DeepSeek’s model cards report Chinese-language benchmarks (for example C-Eval in the DeepSeek-V4.1-Flash card) alongside English ones.

That means there is no official tier list to rely on. If your product targets a specific language — and particularly a lower-resource one — evaluate a sample of real prompts with a native-speaker reviewer before you commit. For background on the model family, see our overview of DeepSeek V4.

Why the model is bilingual-first

DeepSeek’s published technical reports describe a training mix built around English and Chinese, with multilingual data added to broaden coverage. The V3 report’s data section frames the change exactly that way: more mathematical and programming samples, and multilingual coverage expanded “beyond English and Chinese”.

Practically, this suggests three things to plan for:

1. English and Chinese are the languages DeepSeek reports most of its benchmarks in, so they are the most thoroughly documented paths.
2. Other languages work, but DeepSeek publishes no quality figures for them — validate on your own content.
3. Language mixing is a known issue: DeepSeek’s 22 September 2025 Change Log entry for DeepSeek-V3.1-Terminus listed “reduced occurrences of Chinese-English mixing” among fixes for issues reported by users, so instruct the output language explicitly.

Natural language coverage at a glance

Because there are no official per-language ratings, the useful question for each language is what to pin down in the prompt. The table lists the settings that most often matter.

Programming languages

For code, DeepSeek’s lineage includes DeepSeek Coder V2, whose 2024 release notes say it expanded support for programming languages “from 86 to 338”. DeepSeek publishes no per-language list for the current models; both API models are general-purpose and are covered by the same documentation. On which model to pick, DeepSeek’s 10 September 2026 news page says tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, so deepseek-flash is the sensible default.

Fill-in-the-Middle (FIM) completion is available in beta in non-thinking mode only — useful for editor integrations. Tool calling, JSON mode and streaming all work in either thinking or non-thinking mode. For deeper coding workflows, see DeepSeek for coding.

Choosing the right model for multilingual work

Both current models publish open weights under the MIT license. They share a 1M-token context window with output up to 384K tokens and the same OpenAI- and Anthropic-compatible API surface. The differences that matter for language work:

Prices are as of October 2026; off-peak is half the peak rate, and peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Input costs $0.30 per 1M tokens on a cache miss ($0.15 off-peak) on deepseek-flash and $1.32 ($0.66) on deepseek-v4-pro. For full rates and worked budgets, see the DeepSeek API pricing page or the DeepSeek pricing calculator. The legacy IDs deepseek-chat and deepseek-reasoner were scheduled for retirement on 2026-07-24 at 15:59 UTC and are no longer listed in the official docs; migrating is a one-line model= swap, and the base_url does not change.

How thinking mode changes language behaviour

Thinking mode is a request parameter on either model, not a separate model ID, and it is on by default (effort high). You can set it explicitly with extra_body={"thinking": {"type": "enabled"}} plus reasoning_effort="low", "high" or "max", or switch it off with extra_body={"thinking": {"type": "disabled"}}. With thinking on, the response returns reasoning_content alongside the final content.

For multilingual work this matters in two ways. First, DeepSeek does not document which language reasoning_content is written in, so if you display it to users, do not assume it matches the language of the answer. Second, thinking adds latency and output tokens: for a quick chat reply in Spanish, non-thinking mode is usually enough; for a long, high-stakes translation, the default thinking mode may be worth the extra cost — evaluate both on your own content.

Calling the API in another language

Chat requests hit POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com. The API is stateless — you must resend the conversation history with every request. The web chat and mobile app, by contrast, maintain session history for the user. A minimal Python example for a Japanese translation request using the OpenAI SDK, in non-thinking mode so that temperature applies:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key=os.environ["DEEPSEEK_API_KEY"],
)

resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[
        {"role": "system", "content": "You are a professional EN-JA translator. Output only Japanese."},
        {"role": "user", "content": "Translate: 'The quarterly report is attached for your review.'"},
    ],
    temperature=1.3,
    max_tokens=512,
    extra_body={"thinking": {"type": "disabled"}},
)
print(resp.choices[0].message.content)

Three parameter notes for language tasks. temperature follows DeepSeek’s official guidance for non-thinking mode: 0.0 for code, 1.0 for data analysis, 1.3 for general conversation and translation, 1.5 for creative writing; in thinking mode it has no effect. max_tokens should be set generously, and token counts differ by script: DeepSeek’s token-usage guide gives roughly 0.3 tokens per English character and 0.6 tokens per Chinese character, while noting that actual ratios vary. top_p depends on model and mode: on deepseek-flash it only takes effect in thinking mode (and is fixed at 1.0 in the non-thinking example above), while on deepseek-v4-pro it takes effect only in non-thinking mode.

Translation prompting patterns

The single biggest quality lever is often the prompt rather than the model. Five patterns worth using:

1. Specify variant. “Latin American Spanish, neutral register” beats “Spanish.”
2. Pin the audience. “For a B2B SaaS landing page aimed at Berlin-based engineering managers” sets register without you having to enumerate it.
3. Forbid code-switching. “Output only in Polish. Do not include English glosses.” guards against language mixing.
4. Provide a glossary. A short “translate these terms exactly” block fixes brand names and domain vocabulary more directly than any temperature tweak.
5. Round-trip for high-stakes copy. Translate to target, then back to English with a fresh request, and compare. Inexpensive on deepseek-flash, and revealing on tone.

For more on prompt structure, see DeepSeek prompt engineering, and for a translation-specific deep dive, DeepSeek for translation.

Known failure modes by language

• Code-switching to English or Chinese mid-output: DeepSeek itself has shipped fixes for Chinese-English mixing (V3.1-Terminus, September 2025). Guard against it with a “respond only in <language>” instruction in the system prompt.
• Simplified vs Traditional Chinese drift: specify “Traditional Chinese (Taiwan)” or “Traditional Chinese (Hong Kong)” explicitly.
• Honorific register slippage in Japanese/Korean: pin the register (“desu/masu form throughout”).
• RTL display issues: Arabic and Hebrew display problems usually sit in your frontend, not in the response text.
• JSON mode with non-Latin keys: keep keys ASCII; values can be any language. JSON mode is designed to return valid JSON, not guaranteed — include the word “json” and a small example schema in your prompt, and set max_tokens high enough to avoid truncation.

Where DeepSeek fits versus dedicated alternatives

For volume translation between two specific high-resource languages — say, English ↔ German for a documentation pipeline — compare DeepSeek against a dedicated translation API on a sample of your own content before choosing. DeepSeek’s general-purpose models are a natural fit when the task is more than translation: summarising a meeting transcript in one language and producing a localised brief in another, drafting a multilingual support reply that cites the original ticket, or translating code comments while preserving semantics. For a side-by-side with a frontier general-purpose model, see DeepSeek vs Claude, and for the broader landscape, DeepSeek alternatives.

UI language for the chat and app

Separate question: what language is the chat interface itself in? DeepSeek’s official documentation does not list the interface languages of the DeepSeek app and web chat, so check the app’s settings on your device. The interface language and the reply language are separate: the model answers in the language your prompt asks for, so state it explicitly when it matters.

Practical picks by scenario

• English-only product, occasional Spanish/French support tickets: deepseek-flash, non-thinking, temperature 1.3.
• Bilingual zh-en knowledge base: deepseek-flash with a stable system prompt (so repeated prefixes hit the context cache), and the output language fixed per request.
• High-stakes legal or medical translation: thinking mode on, with a native reviewer in the loop; try deepseek-v4-pro only if your own evaluation favours it.
• Multilingual coding assistant: deepseek-flash for editor autocomplete via FIM (non-thinking), and thinking mode for repository-scale tasks.
• Low-resource language research: evaluate first; consider DeepSeek limitations and have a fallback.

For a wider tour of the ecosystem, the DeepSeek beginner guides hub collects the prerequisites — accounts, keys, model picks — in one place.