Last updated: November 2, 2026
AnythingLLM 1.19, released October 26, 2026, lets the self-hosted Docker version run local models inside the container through a built-in engine (AMD’s Lemonade running llama.cpp), so you no longer need a separate Ollama or LM Studio server. It also adds color palettes and dozens of fixes.
Contents
- What’s new at a glance
- Built-in LLM for Docker
- How to turn it on
- Choosing and pulling models
- Color palettes and branding
- Fixes worth knowing
- Power-user workflows
- How it compares
- Frequently asked questions
AnythingLLM, from Mintplex Labs, is an open-source app for chatting with documents, running AI agents and connecting to local or cloud models. This guide focuses on the two October 2026 releases, v1.18.0 on October 13 and v1.19.0 on October 26, and uses only Mintplex Labs’ own release notes, documentation and blog.
What’s new at a glance
| Area | Change | Release |
|---|---|---|
| Docker built-in LLM | Lemonade and llama.cpp run inside the container; off by default | 1.19.0 (Oct 26) |
| Color palettes | Built-in and custom palettes tint the whole UI; Docker admins can enforce one | 1.19.0 |
| Branding settings | Branding and whitelabeling split into Identity, Colors, Sidebar and Browser tabs | 1.19.0 |
| Mistral | Support for Mistral reasoning models with a thinking toggle | 1.19.0 |
| Chat and agents | Notice when a reply hits the max output token limit; workspace model changes apply to agent sessions mid-thread | 1.19.0 |
| Security hardening | Request body size cap for embeds, spreadsheet formulas neutralized in CSV exports | 1.19.0 |
| Reliability | Fixes for scheduled jobs, Outlook, Confluence, Paperless-ngx, AstraDB and document pickers | 1.18.0 (Oct 13) |
Built-in LLM for Docker
The headline feature brings the Desktop app’s built-in engine to the self-hosted image. According to the Built-in LLM for Docker documentation, it is Lemonade, AMD’s open-source local inference server, running llama.cpp on your CPU or GPU. It is available on the latest image and every Docker release after v1.18.0, and it runs on both amd64 and arm64 hosts.
When it is enabled, Built-in becomes the default provider during onboarding of a new instance, with three recommended models labeled Fastest, Balanced and Most capable, sized to your GPU memory or the container’s memory. It works everywhere a provider can be used: the system LLM, workspace and agent LLMs, the chat window model picker and Model Router rules.
The engine picks a backend automatically. The documentation gives this support matrix:
| Host | NVIDIA | AMD or Intel |
|---|---|---|
| Linux | CUDA | Vulkan |
| Windows (Docker Desktop) | CUDA | CPU only |
| macOS (Docker Desktop) | Not applicable | CPU only (no Metal in Docker) |
CUDA requires an NVIDIA GPU from the Turing generation (GTX 16 or RTX 20 series) or newer; older cards fall back to the CPU.
How to turn it on
For an existing instance, the documentation says nothing is reset or migrated, and your current provider stays selected:
- Pull the new image with
docker pull mintplexlabs/anythingllm:latest. - Remove the old container with
docker rm -f anythingllm. Your data stays in the storage folder. - Start it again with your usual command plus
-e ENABLE_BUILT_IN_LLM=true. The setting cannot be turned on from the UI or the.envfile. - Add GPU access:
--gpus allfor NVIDIA (install the NVIDIA Container Toolkit first on Linux), or--device /dev/driplus--group-add $(getent group render | cut -d: -f3)for AMD or Intel on Linux. - Open Settings, then AI Providers, then LLM, select Built-in, and download a model.
The documentation warns that without --group-add the container can see an AMD or Intel GPU but cannot open it, so models silently run on the CPU. On first boot the container downloads the engine (a few MB) into your storage folder, and the first chat downloads the llama.cpp runtime for your hardware, so expect a slower first reply.
Choosing and pulling models
- Model table. Under Settings, AI Providers, LLM, the Built-in table lists each model’s size and whether it supports vision and thinking.
- Faster models. Models marked with a lightning icon use multi-token prediction, which generates several tokens at once.
- Any GGUF repo. Click Pull model from Hugging Face and enter
owner/repoorowner/repo:QUANT, for exampleunsloth/Qwen3-4B-GGUF:Q8_0. Without a quantization,Q4_K_Mis used, and vision models fetch their projector automatically. - Context window. If left empty, AnythingLLM uses the model’s own context length capped at 16,384 tokens to avoid running out of GPU memory. Set a value under Advanced Settings to override it.
The same engine change happened on the desktop. In its October 2026 blog post, Mintplex Labs explains that Desktop’s built-in engine moved from a bundled Ollama to llama.cpp managed by Lemonade, citing faster llama.cpp updates and a smaller install: about 137 MB on a typical Windows PC without an NVIDIA card, compared with about 1.8 GB for the old engine. Existing Desktop users switch by clicking Engine upgrade in the top bar, and models must be downloaded again.
Color palettes and branding
The customization and branding documentation describes color palettes, available since v1.19.0, as a tint for the background, sidebar, chat and settings in both light and dark mode. Open Settings, then Customization, then UI Preferences to pick a built-in palette, including True Dark with pure black backgrounds, or click New palette to build one from a main color and an accent color.
On Docker, admins create palettes on the Colors tab of Branding and Whitelabeling and can turn on Enforce the default palette for everyone.
Fixes worth knowing
The v1.19.0 changelog also includes these changes that affect daily use:
- A notice appears when a response hits the maximum output token limit.
- Changing a workspace’s model now applies to agent sessions already in progress.
- Mistral reasoning models are supported, with a thinking toggle and a deduplicated model list.
- MCP tool results return plain text without the result envelope, and an unsupported MCP ping is no longer fatal.
- Transcription accepts .mov files, and the bulk link scraper parses PDF and file links.
Power-user workflows
- One-container team server. Run a Linux host with an NVIDIA GPU, start the container with
--gpus alland the built-in LLM, and give each workspace its own downloaded model. - Local and cloud mix. Use Model Router rules to send routine prompts to a Built-in model and harder requests to a cloud provider.
- Test a new open model the same day. Pull a GGUF repo from Hugging Face by name and quantization, then compare it in the chat window model picker.
How it compares
Before 1.19, a self-hosted AnythingLLM needed an inference server beside it, such as Ollama or LM Studio; see our Ollama October 2026 guide and LM Studio Bionic guide. That route still works, but the built-in engine removes a second service to install and update. Compare the Open WebUI 0.12 guide. For high-throughput serving, see the vLLM 0.32 guide, and for model downloads, the Hugging Face hf CLI guide.
Frequently asked questions
What is new in AnythingLLM 1.19?
AnythingLLM 1.19.0, released October 26, 2026, adds a built-in local LLM for the Docker image, powered by AMD’s Lemonade server running llama.cpp, plus color palettes for the whole interface and a long list of fixes for connectors, agents, exports and providers.
How do I turn on the built-in LLM in AnythingLLM Docker?
Start or recreate the container with -e ENABLE_BUILT_IN_LLM=true. The documentation says it cannot be turned on from the UI or the .env file, and pulling the new image alone is not enough because the variable is only read when the container starts.
Does the AnythingLLM Docker built-in LLM use my GPU?
Yes, if you pass the GPU into the container. It uses CUDA on NVIDIA GPUs from the Turing generation or newer, Vulkan on AMD and Intel GPUs under Linux, and the CPU otherwise. Docker on macOS cannot use the Apple GPU, so Mac users who want GPU speed should use AnythingLLM Desktop.
Will enabling the built-in LLM change my existing provider?
No. The documentation says your current LLM provider is not changed and nothing is reset or migrated. Built-in is only preselected during onboarding of a new instance; on an existing instance it simply becomes available to select.
Do I still need Ollama with AnythingLLM?
No, the built-in engine can run GGUF models directly. AnythingLLM says connecting to your own Ollama server is still supported and keeps working as before; the change only affects the engine that ships inside the app.
Sources
- AnythingLLM v1.19.0 release notes (GitHub)
- AnythingLLM v1.18.0 release notes (GitHub)
- Built-in LLM for Docker (AnythingLLM Docs)
- Customization and Branding (AnythingLLM Docs)
- Why AnythingLLM is moving from Ollama to Lemonade and llama.cpp (AnythingLLM blog)
Image: “Technician with laptop working on server rack at NERSC” by Derrick Coetzee, Wikimedia Commons, CC0 public domain dedication, cropped and resized. It is a generic photo of a server rack, not an AnythingLLM screenshot. FreeAnswers is not affiliated with Mintplex Labs.
