textgen

Changes

• MTP speculative decoding support: Add draft-mtp as a new --spec-type option. Auto-enabled when loading MTP GGUFs (e.g. Qwen 3.6 MoE MTP builds).
• Web search improvements:
• Add snippet support to the web_search tool: results now include a short text excerpt that often answers the query directly, eliminating the need for a follow-up fetch_webpage call (#7548).
• Drop link URLs from fetch_webpage output (links now appear as plain text instead of [text](url) markdown), significantly reducing tokens used per page.
• Prettier rendering of web_search results in the chat, with a spinner during the call.
• Add an info message to the "Activate web search" checkbox.
• Show live generation speed (tokens/s) and context size while generating (#7563).
• DGX Spark support: Add Linux aarch64 portable builds.
• Electron
• Add "Check for updates" button in the Session tab.
• Add a folder picker for the models directory.
• Add right-click context menu for copying text.
• Add a spellcheck toggle in the Session tab (#7550).
• Store app data in user_data/cache/electron instead of the OS default location.
• Disable DNS-over-HTTPS probes.
• One-click installer: Track the latest release tag instead of bleeding-edge main.
• Auto-detect and auto-select sibling mmproj files when loading a model (#7564).
• Detect mmproj-*.gguf files in the main models folder: They appear in the mmproj dropdown and are hidden from the regular model dropdown.
• Project icon: Add an icon, courtesy of LMLocalizer on Reddit.
• Treat negative --ctx-size values as auto (0).
• UI
• Add drag-and-drop file upload support to the chat input (Gradio fork).
• Reorganize the right sidebar with Mode/Character/Chat style on top.
• Hide reasoning and tools controls in chat mode (only shown in instruct / chat-instruct).
• Fade in new messages, fix scroll-up jump on send.
• Rename "Send dummy message/reply" to "Insert user/assistant message".
• Polish character dropdown in chat tab.
• Tighten spacing between dropdowns and refresh buttons.
• Improve the looks of the Session tab.

Security

• Restrict CORS to localhost by default to prevent drive-by API access. --listen and --public-api opt into network exposure.
• Sanitize character name in load_character to prevent path traversal.
• fix: prevent path traversal in load_template_by_name (#7562). Thanks, @Allen930311.
• UI: Improve web search security by rejecting non-HTTP links.

Bug fixes

• Fix llama-server not being killed when the parent process exits on Windows, e.g. when closing the console window or killing python.exe (#7574).
• Fix streaming output leaking across chats when switching mid-stream (#7555).
• Fix continue-mode regressions across template families.
• Fix incorrect prompts generated with continue mode. Thanks, @MeemeeLab.
• Fix thinking channel being lost across tool-call turns (#7578).
• Fix API model load silently dropping hyphenated arg keys (#7577).
• Fix chat deletion failing when user_data/logs is a symlink (#7579).
• Fix token count not being set in non-streaming mode.
• Keep web search blocks closed when the user closes them mid-stream.
• fix(win): set PYTHONUTF8 for non-ASCII locale Windows compatibility (#7560). Thanks, @jerry78424.
• Set TORCH_VERSION to 2.9.0 to match xformers 0.0.33's torch pin (#7581). Thanks, @AJ-Gazin.

Dependency updates

• Update llama.cpp to https://github.com/ggml-org/llama.cpp/commit/e947228222147356bc7e64154d3439e142481632
• Update ik_llama.cpp to https://github.com/ikawrakow/ik_llama.cpp/commit/40254a51daf485b2b644bcb82a84278d95745ee5
• Update ExLlamaV3 to 0.0.34

Portable builds

TextGen is now a desktop app for local LLMs. Download, unzip, double-click.

> [!NOTE]

> NVIDIA GPU: If nvidia-smi reports CUDA Version >= 13.1, use the cuda13.1 build. Otherwise, use cuda12.4.

>

> ik_llama.cpp is a llama.cpp fork with new quant types. If unsure, use the llama.cpp column.

Windows

| GPU/Platform | llama.cpp | ik_llama.cpp |

|---|---|---|