Settings — Translation

This page configures what translates and how. Translation is off by default and stays a Beta feature — turn it on once you’ve verified your backend works, and start with a small batch before unleashing it on a whole library.

Caution

Local LLM translation quality is improving but still rough at the edges for anime. Cloud backends (DeepL for European languages, Claude/Gemini for cultural nuance) currently produce better results. Test on a sample episode before committing to a backend.

Feature gate

Setting	Default	Values	Effect
Translation enabled	off	toggle	Master switch. When off, every translation control in the UI is hidden or disabled.
Source language	en	ISO 639-1	Language Sublarr translates from.
Target language	de	ISO 639-1	Language Sublarr translates to.
Source language name	English	string	Used in LLM prompts (some prompts work better with full names).
Target language name	German	string	Same.

Backends

Sublarr ships 12 backends. Configure each in the Backends tab; the pipeline picks the active one per translation job. Cost tracking, queue dashboard, per-backend concurrency, and context-windowing are shared across all backends.

Backend	Type	Self-Hosted	API key	Best for
Ollama	LLM	yes	no	Full local control, custom prompts, GPU acceleration.
OpenAI-compatible	LLM	both	yes	GPT-4 endpoints or local OpenAI-shaped servers.
OpenAI ChatGPT	LLM API	no	yes	GPT-4o / GPT-4-turbo via the official endpoint.
Anthropic Claude	LLM API	no	yes	High-quality long-context translation.
Google Gemini	LLM API	no	yes	Fast Gemini 2.x with native multilingual support.
DeepSeek	LLM API	no	yes	Cost-effective, strong on Chinese.
Mistral	LLM API	no	yes	EU-hosted LLM, GDPR-friendly.
DeepL	NMT	no	yes	Best quality for European languages.
Google Translate	NMT	no	yes	Broad coverage, fast, paid per char.
LibreTranslate	NMT	yes	optional	Self-hosted, privacy-focused.
Azure Translator	NMT	no	yes	Enterprise NMT with regional endpoints.
MyMemory	NMT	no	optional	Free tier; useful as zero-cost fallback only.

Per-backend fields

Each backend’s edit panel exposes:

Field	Effect
Enabled	Backend toggle. Disabled backends aren’t tried.
Endpoint URL	Override default endpoint. Useful for self-hosted or proxied installs.
API key / token	Credential. Multi-key pool supported via API Keys.
Model	Which model the backend should use (LLM-only). Pre-populated from the backend’s known good defaults.
Price (per 1M units)	Used for cost tracking. Defaults reflect public pricing at the time the backend was added — verify against your invoice.
Concurrency	How many parallel jobs this backend can serve.
Test	Sends a fixed sample translation; reports success and latency.

Fallback chains

Configure backup backends in case your primary fails. Example:

1. Primary: Ollama (local, fast, free).
2. Fallback 1: DeepL (cloud, high quality, fast).
3. Fallback 2: LibreTranslate (self-hosted backup).

A failed primary triggers automatic re-queue on the next backend in the chain. The job ID stays the same so cost tracking, queue position, and notifications all carry through.

Workers and concurrency

Setting	Default	Values	Effect
Concurrent translations	2	1–8	Per-Sublarr-instance cap on simultaneous jobs.
Batch size (cues per request)	15	5–50	Cues per LLM call. Larger = fewer round trips, but bigger context — risk hitting the model’s limit.
Request timeout (s)	90	30–600	Per-LLM-call timeout.
Max retries	3	1–10	Retries on transient failure before giving up.
Backoff base (s)	5	1–60	First retry waits this; doubles each subsequent.
Temperature	0.3	0.0–2.0	LLM sampling temperature. Lower = more deterministic; higher = more creative.

Series context

When translating an episode, Sublarr can include adjacent cues (and optionally past episodes’ summaries) in the prompt for coherence:

Setting	Default	Values	Effect
Use episode context	on	toggle	Inject N previous + N next cues into the prompt around each batch.
Lookback cues	10	0–50	Number of previous cues.
Lookahead cues	5	0–50	Number of upcoming cues.
Use series context	off	toggle	Carry a brief series summary across episodes (Beta).
Context episodes	2	0–10	How many previous episodes to summarise into context.
Auto-glossary	on	toggle	Inject the matching glossary entries on every job (see Glossary).

Tip

Episode context costs tokens but visibly improves consistency for character names and recurring jokes. Series context (carrying summaries across episodes) is more expensive — leave it off until you’re sure you need it.

Prompt template

Field	Default	Effect
Prompt template	empty (auto-generated)	Override the system prompt. Empty = Sublarr generates from source/target language names. Use only if you know what you’re doing.

Cost & memory

The Cost & Memory tab on this page jumps to the dedicated Cost & Memory view — per-job spend log, translation memory hit rate, and TM management.

Live queue

The Queue tab opens the Translation Queue — live job dashboard with cancellation and retry. Use it to watch jobs in flight and to debug the few that fail.

Configuring Ollama (default backend)

The recommended starting setup if you have a capable host:

1. Install Ollama on the host that runs Sublarr (or a reachable LAN host).
2. Pull a general-purpose model: ollama pull qwen2.5:14b-instruct (or llama3.1:8b-instruct for lighter machines).
3. In Sublarr: Settings → Translation → Backends → Ollama.
4. Set Endpoint URL to your Ollama instance (default http://localhost:11434).
5. Set Model to the pulled model name.
6. Click Test — the response should be the test translation.

Tip

The custom anime-translator-* models we shipped historically are currently broken. Don’t pull them. If a stable replacement lands, this page will name it explicitly.