tylerbcrawford/subgeneratorr v3.3.0

v3.3.0 — Ollama model picker everywhere, engine- and model-aware languages, Customize languages panel, 98 translation targets

on GitHub

This release makes a local stack (Whisper + Ollama) self-sufficient and makes the language lists match what your engine and model can actually do. You can now pick an installed Ollama model for keyterms as well as translation, the language dropdowns follow the selected engine and model, and a new panel lets you trim both lists to the languages you use.

Ollama everywhere

• Ollama for keyterm generation. The keyterm panel gains an "Ollama (Local · Free)" provider, so a fully local stack no longer needs hand-written keyterm lists.
• Installed-model picker. In both the keyterm and Translate panels, the Model dropdown lists the models installed on your Ollama server, with a refresh button and a Host field. If the server can't be reached, it falls back to a free-text model name.
• Thinking models work. Thinking-only models such as qwen3:4b used to spend the whole token budget reasoning and return nothing. Ollama calls now get extra headroom (OLLAMA_THINKING_HEADROOM, default 8192), and inline <think> blocks are stripped.
• Faster qwen3 translation. Replies made of bare strings are now accepted, instead of falling back to slow per-cue requests.

Languages that match your engine and model

• Transcription list follows the Cloud/Local switch. Cloud shows Nova-3's 91 codes, including regional variants and Multi-language. Local shows Whisper's 99 languages.
• 40 Whisper-only languages are now selectable, including Welsh, Swahili, Yoruba, Icelandic, Basque and Malayalam. Each writes its own ISO 639-2 sidecar tag.
• 98 translation targets, up from 58.
• Model-aware targets. Languages the selected model's card doesn't list are greyed out under "Not supported by ", and targets you already picked get a ⚠ chip. Cloud models support everything. Ollama families (Qwen, Gemma, Llama, Mistral, Aya, Command R, Phi, gpt-oss) map to their published lists, and unknown models stay fully selectable with an "unverified" note. The check is advisory: the backend never blocks a translation.

Customize languages

"Customize languages…" at the end of the language dropdowns opens a dialog with Transcription and Translation tabs, search, and select all/none. Your choices are saved on the server to SETTINGS_PATH (default /logs/subgeneratorr-settings.json, on the existing logs volume). Keeping everything selected lets future languages appear automatically.

API

• GET /api/languages and GET /api/languages/translation-support?provider=&model=
• GET/PUT /api/settings/languages
• GET /api/ollama/models?host=
• /api/keyterms/generate accepts provider=ollama with ollama_host.

Upgrading

Pull the new images (:3.3.0 or :latest) and recreate the web and worker containers. No configuration changes are required. Optional new settings: OLLAMA_THINKING_HEADROOM and SETTINGS_PATH. If you use the example compose files, copy in their two new lines from the current examples. The examples forward environment variables one by one, and the v3.3.0 tag's copies don't include these two. To have the Ollama picker list models without typing a host, set OLLAMA_HOST on the web and worker containers.

The full API reference for the new endpoints is in docs/technical.md.

See CHANGELOG.md for the complete list.