AI & ComputeLOT 261119 · 3 min read
Microsoft Open-Sources Aion, the 2.3B-Parameter Model Headed Into Windows
Microsoft open-sourced Aion 1.0 Instruct, a 2.3B-parameter MIT-licensed model that will become the Windows on-device AI by January 2027, replacing Phi Silica.
Key takeaways
- Microsoft published Aion 1.0 Instruct, a 2.3B-parameter model, on Hugging Face on Nov. 18 under the MIT license.
- An optimized Aion plus a Microsoft content moderation model becomes the Windows OS small language model by end of January 2027.
- Aion succeeds Phi Silica, with a 32,768-token context and eight supported languages.
- Microsoft's own tests show 55.15% on MMLU, 74.91% on GSM8K and 83.90% on IFEval.
- The model card flags cross-prompt injection (XPIA) resistance as a safety area to improve.

Microsoft published Aion 1.0 Instruct on Hugging Face on Nov. 18, a 2.3 billion parameter language model that will replace Phi Silica as the Windows on-device AI model by the end of January 2027. The release carries the permissive MIT license, and it came from the Applied Sciences Group inside Microsoft's Windows and Devices division, not from the company's Azure AI organization.
That timing is concrete. The model card says a further optimized version of Aion, paired with Microsoft's content moderation model, becomes the Windows OS small language model by end of January 2027. Microsoft frames the Hugging Face drop as "an introduction to what's coming and a way to maximize transparency with our users," and says quantized versions will follow the unquantized weights already posted.
What Aion actually does
The model targets on-device text work: document comprehension, summarization, rewriting, question answering and chat — all without a constant cloud connection. A 32,768-token context window lets it chew through multi-page documents and long email threads locally, which Microsoft pitches as a privacy play: user data stays on the machine. It supports English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese.
Under the hood it's a 28-layer transformer with roughly 2.29 billion parameters, a 200,029-token vocabulary, and a hybrid attention stack: 21 sliding-window layers plus seven full-attention layers. It runs through Hugging Face Transformers using custom code, with ONNX and GGUF formats listed alongside. A separate ONNX build targets WebGPU.
Aion succeeds Phi Silica, the current Windows on-device model, and its training data extends Phi Silica's recipe: quality-filtered public documents plus synthetic data for math, coding and reasoning. Multilingual data makes up about 6% of the training total. Knowledge cutoff is December 2025.
How does it score?
Microsoft ran its own benchmarks, mostly with EleutherAI's lm-evaluation-harness and GPT-4o as judge where one was needed:
- MMLU (knowledge): 55.15%
- MMLU-Pro (knowledge): 33.35%
- GPQA (knowledge): 28.18%
- GSM8K (math): 74.91%
- MATH-500 (math): 48.00%
- IFEval (instruction following): 83.90%
The knowledge numbers are modest; the instruction-following and grade-school math results are the strong ones. For a model this size, Microsoft's scores are its own, not independently verified.
Where it falls short
The model card is unusually frank about limits. A model this size "can miss details, make misstatements and struggle with multi-step reasoning," the card says, and it "may be susceptible to adversarial attacks." Microsoft names cross-prompt injection (XPIA) among the safety metrics it wants to improve in future releases.
That admission carries weight. Windows is adding AI agents that act on users' behalf, and prompt-injection resistance becomes a real attack surface when local agents read documents and email. Microsoft's answer is architectural: a local content moderation model will run alongside Aion when it ships in Windows, a pairing the company is effectively pre-announcing with this open release.
The open release itself is the strategy. By posting unquantized weights under MIT ahead of the Windows integration, Microsoft lets developers test, fine-tune and probe the exact model family — within its benchmark limits — more than a year before it lands on hundreds of millions of Windows machines. Whether the January 2027 ship date holds, and how the moderation pairing affects what the in-OS model will and won't answer, will define how Windows' local AI stack actually behaves.
mlq.ai (Source)
More from Rebecca Stone
Show full bio
Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,155 articles