Interface, thinking and reply language are three layers

"Changing the language" means at least three layers in DeepSeek Harness: interface, thinking and reply. Editing one does not move the others, and conflating them produces the classic "the menus changed but the answer did not."

Three languages, three meanings

Separate the layers before configuring.

Interface language is the visible shell. It translates copy and does not change model behavior (source).

Thinking language is the internal process. It shapes how reasoning is written and is usually not shown to you.

Reply language is the output. It follows how you ask, the system prompt and model habits.

Interface language: the Russian pack

The pack solves unreadable menus, not unintelligible answers.

1. Install dsh-russian-lang and restart to switch the interface to Russian.
2. It covers core namespaces, translating most menus and notices.
3. Custom overrides let you edit individual strings.
4. Typography and keyboard layout fixes are offered as options.
5. It does not change the model reply language.

Localization is shell translation. It does not teach the model Russian; replies still follow prompts.

Translations can lag. Copy added in a new version may stay untranslated, and seeing English fragments is normal rather than a failed install.

Thinking language: the plugin

It makes reasoning default to Simplified Chinese.

1. It injects a language requirement through a system prompt section.
2. Reasoning defaults to Simplified Chinese while replies are configured separately.
3. Custom prompts can replace the built-in instruction entirely.
4. Compatibility may vary with some modes.
5. A switch must be enabled; it is not always on by default.

Thinking and replying can diverge. You can reason in Chinese and answer in English, because they are different settings (source).

It works at the prompt layer. It is not a native model switch, so long-context compaction can drop it and you must recheck.

Reply language: prompts and model

There is no single language switch for replies; prompts constrain them.

1. State the language explicitly in the system prompt or user message.
2. For multilingual models, explicit instructions beat default behavior.
3. In long sessions the requirement can be compacted away; restate it periodically.
4. Tool output may keep its original language and need a second pass.
5. With mixed-language input, say which language the answer should use.

Reply language drifts most easily. Long conversations, code blocks and tool output all affect it; explicit instructions are steadier than model habits.

Do not change the interface and expect a different answer. Interface and reply language are unbound.

Configure and verify

Set three layers, verify each one.

1. Interface: restart and check whether menus and buttons use the target language.
2. Thinking: trigger a reasoning run and read the language of the process.
3. Reply: ask in the target language and check that answers stay consistent.
4. Mixed input: ask in another language and confirm the reply still follows settings.
5. Long session: recheck after several turns to rule out compaction.

All three verified is the only finished state. Checking the interface alone invites a false sense of completion.

Include a long session in testing. Short sessions passing does not prove stability; compaction and context shifts affect language.

Common misreads

Four misreads dominate.

1. "A Russian pack makes it speak Russian." Interface translation is not model ability.
2. "Thinking language equals reply language." They separate into different settings.
3. "One change lasts forever." Prompt-layer requirements can be compacted or overridden.
4. "Every string gets translated." New copy in new versions may be missing.

All four mistake one layer for the whole. Interface, thinking and reply are separate, and each needs its own check.

Read the settings before judging the effect. Many problems are not failures but the wrong layer configured.

Risk and compatibility

Language plugins carry low risk but still break on upgrades.

1. Version compatibility: interface-injection plugins are sensitive to frontend structure and can fail after an upgrade.
2. Prompt conflicts: several plugins injecting language instructions may overwrite each other.
3. Translation quality: machine-translated copy can mislead, so compare critical actions in English.
4. Data boundary: localization does not upload anything, but still confirm the plugin source.

Install one language plugin at a time and retest after upgrades. Multiple plugins editing interface or prompts are hard to diagnose. Interface plugins fail most often after version changes, so check menus and settings first.

Next steps

Separate the three layers, then configure each. For interface plugins and themes, read How to change themes and wallpapers. For naming variants, read dsh plugin, dsh-plugin and dsh plugins. For prompt mechanics, read Where the system prompt comes from.

Sources: dsh-russian-lang plugin page, dsh-thinking-language plugin page, DeepSeek Harness documentation