You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Local Translator v1.2.0 — Local LLM Auto-Discovery & Headless Auto-Start
🤖 Local LLMs: LM Studio & Ollama Integration (with Auto-Discovery)
Model Auto-Discovery: Automatically scans and lists all downloaded models from LM Studio (~/.lmstudio/models) and Ollama with their exact disk size in GB.
Interactive Model Selector: Settings dialog features a dropdown with all detected models, their parameters, and memory status ([In Memory]).
Headless Background Auto-Start: If LM Studio or Ollama is offline, Local Translator automatically launches the server and loads the selected model into GPU/CPU memory (lms load) without opening heavy Electron windows.
Reasoning Models Support: Built-in intelligent filtration for modern thinking models (Qwen 3, Gemma 4, DeepSeek-R1) stripping <think>...</think> internal blocks to ensure clean, instant translation output.
Auto-Port Configuration: Automatically selects port 1234 for LM Studio and 11434 for Ollama.
Fixed Standalone Runtime: Fixed ctranslate2.dll dependency packaging (cudnn64_9.dll) resolving launch errors on standalone Windows systems.
Русский
🤖 Локальные LLM: Интеграция с LM Studio и Ollama (Автоопределение моделей)
Автоопределение моделей: Автоматическое сканирование каталога моделей LM Studio (~/.lmstudio/models) и сервиса Ollama с отображением точного размера каждой модели на диске в гигабайтах (GB) и параметров квантования.
Интерактивный выбор модели: В окне настроек добавлен выпадающий список всех найденных моделей с кнопками «↺ Обновить список» и «▶ Загрузить в память».
Умный Headless-автозапуск: Если сервер LM Studio или Ollama выключен, при переводе программа автоматически в фоне поднимает локальный сервер и загружает выбранную модель в видеопамять (lms load -y <модель>) без открытия тяжелого графического интерфейса.
Поддержка Reasoning-моделей: Автоматическая очистка блоков рассуждений <think>...</think> для моделей нового поколения (Qwen 3, Gemma 4, DeepSeek-R1) — в результат перевода попадает только чистый текст.
Исправление запуска сборки: Включена необходимая библиотека cudnn64_9.dll для runtime ctranslate2, устранившая ошибку запуска standalone .exe.