we_are_coded.by CODE · The world, decoded
БГ

Microsoft · On-device models

Microsoft's Aion Instruct: a small 2.3 billion parameter model for Windows, under MIT, with eight languages and no Bulgarian

On 18 November Microsoft published the small model Aion Instruct on Hugging Face: 2.3 billion parameters, MIT and a 32K token context, built to run on the device itself. Microsoft plans for an optimized version to become the language model of Windows by the end of January 2027. Bulgarian is not among the eight languages.

Source
Microsoft
Published
The event
Checked on
Responsible editor
Cvetelin Ivanov
How we work
MethodCorrections

Eight languages on the list. No Bulgarian. Aion Instruct is the small language model Microsoft published on Hugging Face on 18 November: 2.3 billion parameters, MIT license, and per the model card it is built to run on the device itself. If you write in Bulgarian, read the list before you get excited.

Fact

The facts: Microsoft published the model Aion 1.0 Instruct on Hugging Face, with 2.3 billion parameters and the MIT license. The timestamp on the page is 18 November 2026. The card describes it as the foundational Small Language Model (SLM) in the Windows ecosystem, optimized for on-device AI, text and operating system use cases. Languages: English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese. The context window extends to 32K tokens. Authors: Applied Sciences Group at Microsoft Windows & Devices. Library names: PyTorch, ONNX, GGUF. The training data is an extension of the data used for Phi Silica, the previous generation Windows SLM, and multilingual data makes up about 6% of the total. Microsoft's plan: a further optimized version of Aion Instruct, coupled with Microsoft's content moderation model, to become the Windows SLM by the end of January 2027. The Hugging Face version is an introduction to what is coming. The unquantized model is published first, and quantized versions are to follow. All of this is from Microsoft's own model card.

Bulgarian is not among the eight. Multilingual data is about 6% of the total, and how much of that goes to which language, the card does not say in the text we read. Microsoft writes that it is working on closing performance gaps across languages and on supporting more languages, and that output quality may differ depending on the language.

Long documents, entirely on the device, in eight languages. Bulgarian is not one of them.

The good part is the context. In Microsoft's words, 32K tokens are enough for multi-page documents, lengthy email threads and complex technical specifications, entirely on-device, with your data remaining private. That is Microsoft's claim. The page has no independent check.

Microsoft is honest about the price of a small size. Such a model can miss fine details, provide misstatements, and fail at complex multi-step reasoning. It was not trained to excel at world knowledge. It may be susceptible to adversarial attacks. XPIA, DIC and ECI are among the metrics to improve in future releases. What exactly they measure and where they stand now, the text we read does not spell out.

One practical thing. The example code on the page for running the model asks for trust_remote_code=True, and the model page carries the custom_code tag.

The test is simple: real Bulgarian text in the Windows version, when it arrives. End of January 2027 is Microsoft's plan. Until then, the only fact about Bulgarian is the list of languages.

The visual is generated code-art, no third-party images.