Release notes

Security

• An update to the server's encrypted connections. The part of the server that encrypts its connections with other machines (HTTPS on port 12443, and the links to remote nodes and to the other servers of a fleet) had security problems, published in October 2026. This version includes the fixed release. Upgrade, especially if your server serves other machines or uses remote nodes.

New

• Signed speech, if you want it. Generated speech already carries metadata saying it is AI-made. Set [provenance] speech = "signed" and WAV and FLAC speech also carries a signed record, in the C2PA format that content tools read, saying that AI created it, which model made it and that OfflinAI Server was the tool. Your server signs it with its own key, made on first start, naming your organisation. Viewers such as a Content Credentials viewer report the signer as "not on their trust list"; that is expected. Business and Enterprise customers can install a certificate from a certificate authority on the C2PA trust list to be shown as a trusted signer. Nothing is sent anywhere: the signing happens on your server and needs no network. Playback starts a little later, because a signature covers the whole file. Speech made on a remote node is marked with metadata but not signed. Guide: section 6.18.
• A "Get release news" link on the console's Overview page. It opens offlinai.com in your browser, where you can sign up for news about new versions. The server itself sends nothing.

Changed: chat applications such as Open WebUI

These changes come from setting up OfflinAI Server with Open WebUI as a private chat assistant for a small office.

• Short model names in the model list. The list that chat applications show their users (GET /v1/models) now names models the short way, for example qwen2.5:1.5b, instead of the longer key. Requests that use the longer key keep working, and an answer repeats the name the request used.
• The model list shows chat models only. Embedding and speech models no longer appear in chat applications' model menus, where they could not chat. Programs that need them can ask for them: GET /v1/models?kind=embeddings, or ?kind=all for every model. The console and GET /v1/models/installed still show everything.
• Answers carry only the standard fields. Chat and completion answers no longer end with an extra block of timing figures. Your usage numbers are unchanged, on the console and in every answer.
• A clearer error for an unknown address. When a program reaches the server under a name it was not told about, it now gets a standard JSON error (403 host_not_allowed) that says which option allows the name, instead of a line of plain text.
• Tidier command-line help. offlinai-server keys --help, usage --help, admin --help and the other commands that do not run models no longer list the model and graphics options. Those options still work where they did.

Not in this release

• Image generation is not included. This version does not create, edit or vary images. The addresses for those requests answer with the standard "not offered" error, so an app that asks can fall back to its own generator, and no image-generation models are listed or offered for download. Everything that reads images stays: chat about images, image descriptions and the screening of images sent for moderation. Speech to text is unchanged.
• Voices of your own are not included either. Text to speech speaks in the four built-in voices, which the model made itself without any person's recording. A recording placed in the server's voices folder is no longer used, and asking for it by name is refused like any unknown voice.

Upgrading

Pull the new image and start the container again the same way: docker rm -f oas, docker pull offlinai/offlinai-server, then docker run -d --name oas -p 127.0.0.1:12436:12436 -v oas-data:/data offlinai/offlinai-server. Settings, keys, models and licences carry over.

If you use Open WebUI or another chat application, the models get new short names in its menu: choose the model again once after upgrading. Chats you already have are kept. If a program of yours reads the model list to find an embedding or speech model, add ?kind=all (or the kind it needs) to its request.

Changed

• The server's address opens the console in a browser. Opening http://127.0.0.1:12436/ in a browser now goes straight to the console at /console/. Before, it showed a single line of plain text meant for programs. Programs that check the address to see whether the server is running get the same answer as before.

Upgrading

Pull the new image and start the container again the same way: docker rm -f oas, docker pull offlinai/offlinai-server, then docker run -d --name oas -p 127.0.0.1:12436:12436 -v oas-data:/data offlinai/offlinai-server. Settings, keys, models and licences carry over.

New

• A model library in the console. The Models page lists about a hundred open-weight models for chat, reasoning, code, vision, embeddings, reranking, moderation, speech and images, each with its size, licence and source. Search it, filter it by kind, by licence (for example only Apache-2.0 and MIT) and by what fits the machine's GPU or memory, and download a model with one click. The list is built into the server, so it works offline and doesn't depend on any outside service. Models whose licence forbids commercial use are not in it. Guide: section 5.2.
• Signed library updates. Once a day the server fetches the list OfflinAI publishes, and uses it only when its signature verifies and it is newer than the built-in one. Turn this off with [library] updates = false, or point [library] url at a mirror for servers without internet access. Nothing is fetched with --offline.
• Search Hugging Face for models that are not in the library: pick a repository, see its GGUF files and their sizes, and download one. A vision model's projector comes with it. If Hugging Face can't be reached, only the search says so.
• The guide in 17 languages. The console shows the administrator and developer guide in the console's language.

Changed

• Removing a model downloaded from Hugging Face also removes its projector and folder when no other model needs them.

Upgrading

Pull the new image and start the container again the same way: docker rm -f oas, docker pull offlinai/offlinai-server, then docker run -d --name oas -p 127.0.0.1:12436:12436 -v oas-data:/data offlinai/offlinai-server. Settings, keys, models and licences carry over.

Fixed

• The container now works with Docker Desktop on Windows and macOS. Before, the container's port 12436 could only be reached with Linux's --network host, so Docker Desktop answered "refused to connect". Start it the same way everywhere: docker run -d --name oas -p 127.0.0.1:12436:12436 -v oas-data:/data offlinai/offlinai-server, then open http://127.0.0.1:12436/console/ on that computer to set the console's password. The computer running the container is treated like the server's own machine; other containers and other machines still need an API key. Guide: section 2.4.

Upgrading

Pull the new image and start the container again with -p 127.0.0.1:12436:12436. A container started with --network host on Linux keeps working. Settings, keys, models and licences carry over.

New

• AI-generated content is marked, as the EU AI Act asks (Article 50(2)). Generated images — previews included — carry IPTC metadata that says they are AI-generated and which model made them; generated speech carries the same in WAV and FLAC metadata; every answer that contains generated text, images or speech carries the AI-Disclosure: mode=machine-generated response header. Response bodies are unchanged, so OpenAI's, Anthropic's and Ollama's SDKs work as before. Prompts are never written into files. Guide: section 6.18.

Changed

• Model names read ollama/… for models from the Ollama registry and hf/… for the Hugging Face models in the catalog (for example ollama/qwen2.5/1.5b, hf/smolvlm2/2.2b). The names older clients use for the same models keep working everywhere — in requests, key scopes, project policies, model states and pins — and an answer repeats a model's name as the request spelled it. Usage recorded before this release keeps the names it was recorded with.
• Compatibility mode for older clients: ask for it with X-API-Compat: legacy (the values earlier releases accepted still work). In this mode, model listings use the names older clients look models up by. Guide: section 6.15.
• WAV speech now has metadata chunks before the audio data. Programs that read WAV files normally are unaffected; a program that assumes the samples start at byte 44 must find the data chunk instead.
• Remote workers: update nodes together with the server, so both use the same model names.

Upgrading

Nothing to change: install the new version over the old one, or pull the new image. Settings, keys, models and licences carry over.

The first release: OfflinAI Server for Windows and Linux, as a Windows service, Debian and RPM packages, a container image and a Helm chart, with the web console and its built-in guide, and licensing that works online or air-gapped.