New Features

• Reasoning effort controls - pick how hard the model thinks per chat for Anthropic, OpenAI, Gemini, DeepSeek, Ollama, and GenieX Engine models
• Pull a model from Hugging Face straight into AnythingLLM with a deep link (anythingllm://pull-hf?repo=unsloth/gemma-3-1b-it-GGUF&file=gemma-3-1b-it-Q4_K_M.gguf)
• Tinyscribe 0.1.0 with new speaker UI controls and improved performance
• Personalized greetings on the home screen and empty chats that change with the time of day
• Gemini is now available as an image generation provider
• llmman is now available as an image generation provider
• Firecrawl is now available as an agent web search provider

Improvements

• Onboarding updated to cover magic features, with a new privacy UI
• Workspace settings now save automatically as you edit, with no more "Update workspace" button
• Provider search menus now autofocus and support arrow keys, Enter, and Escape
• Workspace temperature now applies to agents, and is left out of the request entirely when unset so the provider uses its own default
• Temperature is no longer sent to Anthropic models (Anthropic and Bedrock), since newer Claude models reject it
• OpenAI GPT models on AWS Bedrock now work in chat and agents, with native tool calling
• Generic OpenAI max tokens can be set to 0 to leave max_tokens out of the request
• TogetherAI now has a configurable max tokens setting to prevent truncated responses
• SearchApi web search now returns results for all supported engines, including Google Jobs and Google Maps
• Text uploads in UTF-16 and other non-UTF-8 encodings are now read correctly, falling back to UTF-8
• Spreadsheet dates and times are now read as dates instead of serial numbers
• Generated images are now included in chat exports
• Text-to-speech now skips emoji, Markdown, and thinking blocks for cloud voices, Telegram voice replies, and the desktop assistant
• GenieX bumped to 0.7.1 with token metric reporting

Bug Fixes

• Fixed uploaded images being lost when switching from chat to @agent mid-session
• Fixed developer API document attachments being ignored in agent chats
• Fixed Anthropic replies coming back empty when the first content block was not text
• Fixed compressed RAG context being dropped when the prompt or a chunk contained "Context:"
• Fixed Weaviate and Astra DB similarity scores so unrelated chunks no longer pass the similarity threshold
• Fixed Milvus and Zilliz failing to embed into workspaces whose slug starts with a digit
• Fixed Azure embedding deployment changes not resetting vectors or showing a warning
• Fixed malformed agent WebSocket messages crashing the server
• Fixed the Ollama and llmman Keep Alive setting being ignored in agent chats
• Fixed a workspace still using its old router after switching back to the system default provider
• Fixed embeds ignoring their message history limit when the model router is on
• Fixed embeds ignoring the workspace's custom refusal message when the workspace has no documents
• Fixed embeds saving empty replies that counted toward chat limits
• Fixed multi-sheet XLSX uploads losing sheets with non-Latin or similar names
• Fixed HTML-only mbox messages with attachments being skipped
• Fixed every message in an mbox upload getting the same title
• Fixed API uploads with addToWorkspaces only embedding the first sheet or message of a file
• Fixed created Excel files turning IP addresses, version numbers, and decimal commas into wrong numbers
• Fixed Confluence code blocks losing their code on import
• Fixed deep website crawls resolving relative links against the start URL
• Fixed links whose URL path cannot be percent-decoded failing to save
• Fixed DuckDuckGo result links being double-decoded and titles containing HTML
• Fixed chat exports including deleted, reset, and API session chats
• Fixed JSONL and Alpaca chat exports using the default prompt and merging embed chats across workspaces
• Fixed scheduled jobs running a day early or late when local time falls on a different UTC date
• Fixed the Google Calendar agent skill ignoring multi-user mode and disabled sub-skills
• Fixed new agent flow blocks sharing an id after a block was deleted
• Fixed filesystem-edit-file expanding $ patterns in new text
• Fixed Telegram bot commands running when a link contained /new, /reset, or /model
• Fixed Telegram link URLs breaking when they contained Markdown characters
• Fixed Telegram replies not rendering c++, ~~~, and nested fenced code blocks as code
• Fixed Kokoro voices not loading when the base URL had a path or trailing slash
• Fixed Fireworks models and Kokoro voices not loading with a saved API key
• Fixed the LiteLLM API key not being saved or masked
• Fixed manually set .env values being dropped when settings were saved
• Fixed turning off password protection not surviving a restart
• Fixed users and admins being unable to clear a bio
• Fixed mobile sidebar drawers being clipped at the bottom
• Fixed AnythingLLM Mobile clearing the default thread when it sent a deleted thread

Pinned Download Links

Revision 1.17.0: