Microsoft Aion 1.0 Instruct: The Small Language Model Headed to Windows
Microsoft released Aion 1.0 Instruct, a 2.3B-parameter on-device model set to become the Windows OS SLM by January 2027. Here's what's inside.

What is Microsoft Aion 1.0 Instruct?
Aion 1.0 Instruct is a 2.3 billion parameter small language model that Microsoft’s Windows Applied Sciences Group released as an open-weight model on Hugging Face. It’s built for on-device text tasks like summarization, rewriting, and chat, and it’s designed to run locally on Windows hardware rather than calling out to the cloud. Microsoft has said a further optimized version, paired with a separate content moderation model, will become the official Windows OS SLM by the end of January 2027. This public release is explicitly framed as a preview.
TL;DR
- Aion 1.0 Instruct is a 2.3B-parameter instruction-tuned model from Microsoft’s Windows Applied Sciences Group, released under the MIT license with PyTorch, ONNX, and GGUF support.
- The model is the successor to Phi Silica, the small language model that currently powers on-device Windows AI features, and its training data builds directly on top of Phi Silica’s dataset.
- Microsoft plans to ship a more optimized version of Aion Instruct, combined with a local content moderation model, as the general-availability Windows OS SLM by the end of January 2027.
- The model supports a 32,768-token context window and eight languages, with English dominating the training mix and multilingual data making up only about 6%.
- Benchmark scores show solid instruction-following (IFEval 83.90) but modest world knowledge (MMLU 55.15, GPQA 28.18), which lines up with Microsoft’s own description of it as weak on broad factual recall.
- Microsoft is publishing the unquantized weights first, with quantized versions to follow, and the card is explicit that this release exists to gather feedback before the Windows integration ships.
- The architecture uses a hybrid attention pattern (mostly sliding-window layers with a handful of full-attention layers) plus grouped-query attention, a common recipe for keeping small models fast on constrained hardware.
Why is Microsoft putting a language model into Windows itself?
Microsoft has been building toward on-device AI in Windows for a while, and Phi Silica was the first concrete step: a small model baked into the OS to handle tasks like rewriting text or summarizing content without sending data to a server. Aion Instruct is positioned as the next generation of that same idea. The model card calls Aion Instruct “the foundational Small Language Model (SLM) in the Windows ecosystem,” which signals that Microsoft wants a model living locally on hundreds of millions of PCs, handling lightweight language tasks instantly and privately.
The pitch for doing this on-device rather than in the cloud is straightforward: lower latency, no dependency on a network connection, and better privacy because user text doesn’t have to leave the machine. A 2.3B-parameter model is small enough to run on consumer hardware without requiring a dedicated GPU cluster, which is the whole point of an OS-level SLM. It won’t replace a frontier cloud model for hard reasoning, but it can handle everyday text work like summarizing an email thread or rewriting a paragraph fast and locally.
What are the technical specs of Aion 1.0 Instruct?
The model card lays out a fairly detailed architecture. Some of the key numbers:
- Parameters: approximately 2.29 billion, with tied input and output embeddings
- Layers: 28, with a hybrid attention setup of 21 sliding-window layers (window size 2048) and 7 full-attention layers
- Attention mechanism: grouped-query attention (GQA) with 24 query heads and 8 key/value heads, head dimension 128
- Positional encoding: RoPE with YaRN scaling applied to the full-attention layers, which is how the model stretches to its full context length
- Context length: 32,768 tokens
- Vocabulary size: 200,029 tokens
- Precision: bfloat16
- Knowledge cutoff: December 2025
The hybrid sliding-window/full-attention design is a common trick for keeping memory and compute costs down in small models while still allowing some layers to see the entire context. Combined with GQA, which reduces the memory footprint of the attention mechanism, this is an architecture tuned for efficiency on consumer devices rather than raw scale.
The model also supports tool calling through a defined chat template, with tool schemas injected into the system prompt and tool calls emitted in a structured <|tool_call|> format. That matters for anyone who wants to wire Aion Instruct into an agent-style workflow where it needs to invoke functions rather than just generate text.
How does Aion Instruct perform on benchmarks?
Microsoft published scores across knowledge, math, instruction-following, and translation benchmarks, run largely through EleutherAI’s lm-evaluation-harness. The results tell a consistent story: this is a model built for following instructions and handling structured text tasks, not for deep world knowledge or complex reasoning.
On knowledge benchmarks, Aion Instruct scored 55.15 on MMLU, 33.35 on the harder MMLU-Pro, and 28.18 on GPQA, a graduate-level science QA benchmark. Those numbers are modest, and Microsoft’s own documentation acknowledges this directly, stating the model “was not trained to excel at world knowledge.”
Build an onboarding app for 4,000 new hires.
Build ↵On math, it posted 74.91 on GSM8K (grade-school math word problems) but dropped to 49.56 on the more adversarial GSMPlus and 48.00 on MATH-500, suggesting it handles straightforward arithmetic reasoning better than multi-step or harder problems.
Instruction-following is where the model looks strongest: 83.90 on IFEval, which measures how well a model follows explicit formatting and content instructions. That’s the capability that matters most for an OS-level assistant that needs to reliably do what it’s told (rewrite this, summarize that, format this way) rather than reason from first principles.
Translation performance, measured by chrF++, came in at 43.18, and multilingual MMLU landed at 45.51, both noticeably behind the English-only MMLU score. That gap tracks with the training data: multilingual content makes up only about 6% of the overall training mix, even though the model officially supports English, German, Spanish, Portuguese, Italian, French, Chinese, and Japanese.
What are the limitations and safety considerations?
Microsoft is unusually candid in the model card about what a 2.3B model can’t do. It states plainly that small models like Aion Instruct “can miss fine details, provide misstatements, or show inability to follow complex multi-step reasoning,” and that it lacks “the complex reasoning needed to understand all the nuances of safety, including attempts to induce unsafe behavior.” The card flags the model as potentially susceptible to adversarial prompting.
This is why Microsoft isn’t shipping Aion Instruct alone into Windows. The plan calls for pairing it with a separate local content moderation model when it reaches general availability, acting as a guardrail layer since the core model itself isn’t considered reliable enough to police its own output. The card also names three safety-related metrics, XPIA, DIC, and ECI, as areas it intends to improve in future releases, without detailing current scores for them in the text.
The out-of-scope guidance is standard but worth noting for anyone building on top of it: Microsoft explicitly says the model shouldn’t be used for automated decision-making in high-risk contexts like medical or legal judgments, and developers are expected to evaluate accuracy, safety, and fairness themselves before deploying it for a specific use case.
Is Aion Instruct worth using right now?
For most developers, the honest answer is that this release is a preview, not a product. Microsoft says explicitly that the Hugging Face version is meant “to be an introduction to what’s coming and a way to maximize transparency with our users, as well as welcome their feedback.” The unquantized weights are out first, with quantized versions to follow, which suggests the current release isn’t yet optimized for the lightweight, low-footprint deployment that an OS-embedded model will eventually need.
If you’re experimenting with small on-device models, building Windows-adjacent tooling, or just want to see where Microsoft’s SLM strategy is headed after Phi Silica, it’s worth downloading and testing. If you need a production-ready small model today for summarization or rewriting tasks, there are more mature open alternatives in a similar size class already optimized for deployment. Aion Instruct’s real significance is less about its current benchmark scores and more about the fact that it previews the model that will sit inside Windows itself by early 2027.
Frequently Asked Questions
What is Microsoft Aion 1.0 Instruct?

Nothing gets lost.
It’s a 2.3 billion parameter instruction-tuned small language model from Microsoft’s Windows Applied Sciences Group, released as open weights for on-device text tasks like summarization, rewriting, and chat.
Is Aion Instruct replacing Phi Silica?
Yes. The training data for Aion Instruct extends the dataset used for Phi Silica, the current Windows on-device SLM, and Microsoft describes Aion Instruct as the next-generation foundational model for the Windows ecosystem.
When will Aion Instruct be built into Windows?
Microsoft states that an optimized version of Aion Instruct, combined with a local content moderation model, will become the Windows OS SLM with general availability by the end of January 2027.
What is the context window and language support?
Aion Instruct supports a 32,768-token context window and eight languages: English, German, Spanish, Portuguese, Italian, French, Chinese, and Japanese. English dominates the training data, with multilingual content making up roughly 6% of the total.
What license is Aion Instruct released under?
It’s released under the MIT license, with model files available in PyTorch, ONNX, and GGUF formats.




