Microsoft published Aion-1.0-Instruct-2.3B, a 2.3-billion-parameter language model for on-device use in Windows, on Hugging Face on Nov. 18, 2026 under an MIT license.

According to the model card1, the release is a preview. A further optimized version, paired with Microsoft's content moderation model, will become the Windows OS Small Language Model by the end of January 2027. Microsoft said the Hugging Face version is meant to be "an introduction to what's coming" and a way to invite user feedback.

Background

The model card says the training data is "an extension of the data used for training Phi Silica," which it describes as the previous-generation Windows OS small language model. The data includes filtered public documents and synthetic data for teaching math, coding and common-sense reasoning. Multilingual data makes up about 6% of the total, and the card says output quality may differ by language. The model supports English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese.

The weights ship as bfloat16 safetensors. Running the model requires trust_remote_code=True because it uses custom modeling code stored in the repository. The card lists a hybrid attention design with a 32,768-token context window, reached with YaRN scaling at a factor of 16 on the full-attention layers.

Scores are Microsoft's own

The card reports that the model scored 74.91 on GSM8K1, with the note "loose matching," and 45.51 on MMMLU, a multilingual version of MMLU. On instruction following it lists 83.90 on IFEval (the mean of prompt-level strict and instruction-level strict), 32.00 on IFBench (prompt-level loose) and 64.70 on Multi-IF.

Microsoft said it ran the benchmarks with EleutherAI's lm-evaluation-harness where a task implementation existed, and used GPT-4o as the judge for tasks that need one. The card includes no independent evaluation.

Stated limits and out-of-scope uses

The card says the model "was not trained to excel at world knowledge." It also states: "small language models such as Aion Instruct can miss fine details, provide misstatements, or show inability to follow complex multi-step reasoning."

On safety, the card says the model "does not have the complex reasoning needed to understand all the nuances of safety, including attempts to induce unsafe behavior," and "may be susceptible to adversarial attacks." It names three safety metrics, XPIA, DIC and ECI, as "amongst the key metrics to improve in future releases." Microsoft said a content moderation model will run alongside Aion Instruct when it reaches general availability in Windows.

The card says the model should not be used to generate harmful, violent, explicit or offensive content, to create misleading content such as fake news, or for automated decision-making in critical contexts such as medical diagnoses or legal judgments.

Quantized and related builds

The card says the unquantized model is published first, "to be followed by quantized versions." About three minutes after the main repository, Microsoft posted an ONNX WebGPU version2. Its weights are 4-bit symmetric block-wise, exported from a quantization-aware-trained checkpoint, with FP16 activations and key-value cache. The quantized weights total about 1.3 GB. The build runs on any ONNX Runtime build that provides the WebGPU execution provider, including in a browser. Its default maximum length is 4,096 tokens, which can be raised to the model's 32,768-token context at the cost of a larger pre-allocated cache.

The ONNX card says quality relative to the bfloat16 model "will be reported here," so no quantized-model results are published yet. It warns that "quantization may introduce additional quality degradation."

The same card lists a GGUF build for llama.cpp and an ONNX build for NPUs under the asgmodels account on Hugging Face. The card does not say who operates that account.

A name shared with another product

A separate hosted service from AionLabs, Aion 3.5 Mini, shares the name, according to an orcarouter blog post3 comparing the two. The post says it draws its Microsoft details from the model card.