Albireo
Your private AI assistant, running on your own computer
Albireo runs language models directly on your Windows 11 computer. Your conversations, documents and models stay on the machine: no account to create, nothing sent outside.
- Free
- Open source
- Windows 11, x64
- Works offline
- Version 1.0.23
- Released 15 November 2026
- License Boost Software License 1.0

- 56.8
- tokens per second on a laptop GPU, against 23.6 on its processor
- 22
- models in the signed catalog, from 7 publishers
- 5
- engines: processors, Intel NPU and GPU, NVIDIA GPU
- 0
- account, subscription or data sent outside
The strengths of Albireo
Private by design
The service listens on the local address only. A question, a document or an image leaves the computer only if you switch the relay on.
Fast on the hardware you have
Its engine runs quantized models with integer kernels and takes advantage of the Intel NPU and GPU, or of an NVIDIA card, when the computer has one.
Up and running in minutes
The complete installer ships with its models: Albireo answers as soon as the installation ends, even without a connection.
Compatible with your tools
The local interface follows the OpenAI interface: your client libraries, editors and scripts work unchanged.
Every function of an assistant, running locally
- Chat
- Discussions kept in the browser, an incognito mode, a search across the discussions, and the reasoning of the models that reason shown while they think.
- Documents
- Word, Excel, PowerPoint, OpenDocument, PDF, RTF, HTML, text, CSV, Markdown and source code, read in the browser and attached to a message or a project.
- Images
- The models that read images take PNG and JPEG pictures, attached, pasted or dropped into a question.
- Projects
- Discussions grouped with instructions and documents that form their shared context, searched by meaning and by words.
- Agent mode
- The model calls tools that run in the page: calculator, sandboxed JavaScript, reading and search of the documents, up to 100 steps per answer.
- Web access
- An optional search of Wikipedia and the reading of public pages and PDF files, off in each new discussion.
- Reading aloud
- Each answer can be read aloud by the speech synthesis that comes with Albireo, without network.
- An interface for other programs
- Chat completions, completions, embeddings, speech and models follow the OpenAI interface, with streaming and images, documented in a Swagger page.
- Offline mode
- On a computer without network, the service opens no connection: the catalog, the models and the updates come from a signed bundle brought on a disk.
- Relay to an external engine
- A discussion can be relayed to an external engine: any interface compatible with OpenAI, the cloud of Ollama or Claude. Its texts are pseudonymized by Pseudalis before they leave the computer, and an engine of the computer or of the local network remains usable offline.
The model catalog of Albireo
Albireo runs nearly any open-weight language model published in the GGUF format, in both of today's main structures: dense models, where every parameter works for each token, and mixture-of-experts models, where a router activates a few experts. The catalog below lists the 22 models tested and offered for download, each checked by its hash.
Dense models
Architectures llama, mistral, mistral3, granite, phi3, qwen2, qwen3 and qwen3vl.
Mixture-of-experts models
Architectures mixtral, granitemoe, qwen3moe and qwen3vlmoe.
| Model | Publisher | Functions | Reads | Size | License | Official source |
|---|---|---|---|---|---|---|
| Qwen3-VL 4B Instruct default | Alibaba | general, code, security | text, images | 2.95 GB | Apache-2.0 | Hugging Face |
| Qwen3-VL 4B Thinking | Alibaba | general, code, security | text, images | 2.95 GB | Apache-2.0 | Hugging Face |
| Qwen3-VL 2B Thinking | Alibaba | general, code, security | text, images | 1.55 GB | Apache-2.0 | Hugging Face |
| Qwen3 Embedding 0.6B default | Alibaba | document search | text | 639.2 MB | Apache-2.0 | Hugging Face |
| Qwen3-VL 2B Instruct | Alibaba | general | text, images | 1.55 GB | Apache-2.0 | Hugging Face |
| Qwen3-VL 8B Instruct | Alibaba | general, code, security | text, images | 5.78 GB | Apache-2.0 | Hugging Face |
| Qwen3-VL 8B Thinking | Alibaba | general, code, security | text, images | 5.78 GB | Apache-2.0 | Hugging Face |
| EuroVLM 9B Preview | EuroLLM consortium | general | text, images | 6.42 GB | Apache-2.0 | Hugging Face |
| Phi-4-mini | Microsoft | general | text | 2.49 GB | MIT | Hugging Face |
| Ministral 3 3B | Mistral AI | general | text, images | 2.99 GB | Apache-2.0 | Hugging Face |
| Ministral 3 8B | Mistral AI | general | text, images | 6.06 GB | Apache-2.0 | Hugging Face |
| EuroLLM 9B Instruct | EuroLLM consortium | general | text | 5.58 GB | Apache-2.0 | Hugging Face |
| Granite 4.2 3B | IBM | general, code | text | 2.24 GB | Apache-2.0 | Hugging Face |
| Granite 4.2 8B | IBM | general, code | text | 5.35 GB | Apache-2.0 | Hugging Face |
| Qwen3-VL 30B A3B Thinking | Alibaba | general, code, security | text, images | 19.27 GB | Apache-2.0 | Hugging Face |
| Mistral Small 3.2 24B | Mistral AI | general | text, images | 15.21 GB | Apache-2.0 | Hugging Face |
| Qwen2.5 Coder 1.5B | Alibaba | code | text | 1.12 GB | Apache-2.0 | Hugging Face |
| Qwen2.5 Coder 7B | Alibaba | code | text | 4.68 GB | Apache-2.0 | Hugging Face |
| Devstral Small 2 | Mistral AI | code | text, images | 15.21 GB | Apache-2.0 | Hugging Face |
| CyberSecQwen 4B | athena129 (lablab.ai) | security | text | 2.50 GB | Apache-2.0 | Hugging Face |
| Foundation-Sec 1.1 8B | Cisco | security | text | 4.92 GB | Llama 3.1 Community License | Hugging Face |
| Foundation-Sec 8B Reasoning | Cisco | security | text | 4.92 GB | Llama 3.1 Community License | Hugging Face |
Sizes count the decoder and its image encoder. Each model keeps the license of its publisher.
Speed on everyday hardware: the techniques of Albireo
Albireo runs models with its own engine, designed for the processors and graphics cards already in service. Each technique below is measured in the README of the product.
Quantized kernels
The weights stay in the memory mapping of the GGUF file, in their quantized form. Integer kernels compute 32 products where a vector of floats computes 8.
Reading a prompt: 216 tokens per second for Qwen3-VL 4B and more than 1,300 for Qwen3 0.6B (Intel Core Ultra 9 275HX, 24 threads); on two cores, 230 against 126 on floats (Qwen3 0.6B, AVX-VNNI)
Drafted tokens
An answer writes several tokens per step when its context foretells them, a form of speculative decoding. The model keeps the tokens it would have drawn itself, so that the answers keep the distribution of the sampling.
25.5 against 19.7 tokens per second when rewriting code (Qwen3 0.6B, two cores); 26.3 against 23.6 on 24 threads (Qwen3-VL 4B)
An engine for each kind of processor
Portable, AVX2 and AVX-VNNI engines for processors, the NPU and the integrated GPU of Intel through OpenVINO, NVIDIA graphics processors through CUDA. The service measures them on the machine and runs the fastest.
Writing: 80.6 to 87.7 tokens per second for Qwen3 0.6B with the processor engine (Intel Core Ultra 9 275HX, 24 threads)
Layers computed on the NVIDIA card
The CUDA engine keeps the weights on the card in their GGUF formats and computes the layers of a dense model there, attention included for the generated tokens.
56.8 against 23.6 tokens per second (Qwen3-VL 4B, RTX 5070 Ti Laptop GPU against 24 threads)
V²RAM extended memory
The extended device memory of the library lets the copies of the weights go beyond the memory of the card, into host memory or a file, and brings them back ahead of their use.
Compact conversation memory
Keys and values are kept in 8 bits, in pages shared between the conversations that start with the same text.
34 bytes against 64 per group of values, with the same next token on 210 positions
Several conversations at once
The requests of every client are computed together in each forward pass, and the keys and values of a finished request stay cached for the next turn.
32 to 40% less time per step with four answers written together
An OpenVINO graph on the Intel processor and GPU
With OpenVINO, a whole model is computed as a single graph on the processor or on the integrated GPU of Intel, and the NPU takes matrix products. The graph reads the first pass of a prompt faster; the processor engine stays faster on the following passes and when writing, and the measure of the engines, started from the tray icon, keeps the best choice for the machine.
First reading of a prompt: 260 tokens per second with the graph against 216 with the processor engine (Qwen3-VL 4B, Intel Core Ultra 9 275HX, 24 threads)
| Engine | Runs on |
|---|---|
cpu | Every x86-64 processor |
cpu-avx2 | Processors with AVX2, FMA and F16C: Intel Haswell, AMD Zen and later |
cpu-avxvnni | Processors with AVX-VNNI: Intel Core from Alder Lake, Intel Xeon from Sapphire Rapids, AMD Zen 5 |
openvino | NPU and integrated GPU of Intel processors, through OpenVINO |
cuda | NVIDIA graphics processors, through CUDA |
Screenshots








Audiences and uses of Albireo
Teams whose data must stay in-house
Legal, health, finance, public bodies and industry: questions about internal documents stay inside the organisation.
Developers
A local OpenAI-compatible endpoint for editors, scripts and tests, with code models in the catalog.
Security teams
Models specialised in cybersecurity, an offline mode and a signed bundle for isolated computers.
Integrators and software publishers
A component to embed or to distribute under their own brand, under a permissive licence.
Characteristics
| Platform | Windows 11, x64 |
|---|---|
| Programs | Inference engines, Windows service, web application, tray icon and installer |
| Model format | Quantized GGUF: Q4_0 to Q8_0, Q2_K to Q6_K and IQ4_NL |
| Architectures | llama, mistral, mistral3, granite, granitemoe, phi3, qwen2, qwen3, qwen3moe, qwen3vl, qwen3vlmoe, mixtral |
| Image encoders | Qwen3-VL, LLaVA-NeXT (EuroVLM) and Pixtral (Ministral 3, Devstral Small 2, Mistral Small 3.2) |
| Models that come with Albireo | Qwen3-VL-4B-Instruct with its image encoder (2.95 GB) and Qwen3-Embedding-0.6B (0.64 GB) |
| Catalog | 22 models from 7 publishers, signed and checked against their hashes |
| Context | The context window each model was trained for |
| Memory | 8 GB for the model installed by default |
| Local interface | https://localhost:8443, compatible with the OpenAI interface, documented with Swagger |
| Conversation formats | ChatML, Mistral, Llama 3, Granite, Phi-4-mini and Foundation-Sec, written natively |
| Interface languages | English, French, German, Italian and Spanish |
| Updates | Signed feed, installers checked by size and SHA-256 |
| License | Boost Software License 1.0, Community Edition |
Security you can audit
- Local only
- The service listens on 127.0.0.1 and refuses the requests whose host or origin is not local.
- A certificate authority of the computer
- The HTTPS certificate comes from an authority made on the computer with an ECDSA P-256 key, restricted by name constraints to localhost.
- Keys protected by Windows
- The private key of the server and the key of the relay are encrypted with DPAPI and never written in clear.
- A signed supply chain
- The update feed and the model catalog are signed with ECDSA P-256, and every downloaded file is checked against its SHA-256.
- An open interface
- The routes follow the OpenAI interface and are described with OpenAPI, so that existing client libraries work unchanged.
- Isolated speech synthesis
- The synthesizer runs in a separate process of low integrity, without privileges and bounded in memory.
Good to know
- Price and licence
- Albireo Community is free, for personal and professional use, under the Boost Software License 1.0. Each model keeps the licence of its publisher.
- Hardware
- A Windows 11 x64 computer with 8 GB of memory runs the default model. A graphics card or an Intel NPU speeds it up and is not required.
- Without a network
- Everything works offline once the models are installed. An isolated computer receives the catalog, the models and the updates from a signed bundle.
- Your data
- Conversations are kept in the browser of the computer. The service reaches the outside only for updates, for the models you download, and for the web access or the relay you switch on.
- Updates
- The service reads a signed feed and verifies each installer by its size and its SHA-256 before installing it.
Installation
Each version comes with two installers. The complete installer ships with the models supplied with Albireo: the assistant answers as soon as the installation ends, without a connection. The online installer holds the same programs without the models: the service downloads the default models at its first start.
The wizard asks for the language, the license, the folders and the options. The command line covers silent deployments:
albireo-setup /silent [/dir <folder>] [/data-dir <folder>] [/models <folder>] [/port <port>] [/offline]
albireo-setup /uninstall [/silent] [/keep-data]
albireo-setup /models-dir <folder>
albireo-setup /service start|stop|restart
albireo-setup /lang en|fr|de|it|esWith Pseudalis on the same computer
The relay can pseudonymize every text before it is sent and restore the values in the answers as they stream. The table of substitutes stays in the memory of the service, and images never leave the computer: the local model describes them.
Install Albireo and ask your first question
The complete installer ships with its models: Albireo answers as soon as the installation ends.