v4.1.1
The popular local LLM interface now lets models execute Python functions like web search and calculations during chat.
Oobabooga has launched version 4.1.1 of text-generation-webui, bringing sophisticated agent capabilities to local AI setups. The headline feature is native tool-calling: models can now execute custom Python functions during conversations. Users simply drop .py files into a tools directory, with five examples provided including web_search, calculate, and get_datetime. During streaming, tool calls appear as collapsible accordions showing function arguments and outputs. The update also replaces html2text with trafilatura for cleaner webpage text extraction, reducing boilerplate and saving tokens in agentic loops.
Significant OpenAI API compliance improvements make the interface more developer-friendly. The update adds full logprobs support across llama.cpp, ExLlamaV3, and Transformers backends, plus proper tool_calls response formatting and stream_options support. Performance enhancements include optimized chat streaming (DOM updates once per animation frame), faster startup by removing unnecessary imports, and increased context window support up to 1M tokens via UI sliders. The release also introduces 'incognito chat' for temporary conversations, refactors reasoning extraction into a standalone module supporting multiple model formats, and removes deprecated rope scaling parameters for modern long-context models.
- Native tool-calling lets models execute Python functions like web_search and calculate during chat, with visual accordion displays
- Full OpenAI API compliance with logprobs support, proper tool_calls formatting, and stream_options for all major backends
- Performance upgrades: 1M token context support, faster startup by 0.5-0.8 seconds, and optimized streaming updates
Why It Matters
Enables sophisticated agent workflows on local hardware, making advanced AI capabilities accessible without cloud dependencies.