Sign in
Pending benchmarksTracking since Nov 20, 2026

Is Aion 1.0 Instruct state of the art?Pending benchmarks.Announced Nov 19, 2026; no independent results yet.

Microsoft Windows Applied Sciences Group's instruction-tuned on-device small language model (2.3B parameters, 32K context, eight languages). Official Microsoft Edge blog, published November 19, 2026: generally available as ungated MIT-licensed open weights on Hugging Face, and now Edge's built-in small language model for the Prompt API origin trial, replacing Phi-4-mini in Edge. No official cloud API price is stated; it runs locally.

Announced facts

The details so far

Lab
Microsoft
Announced
Nov 19, 2026
API
Not yet
Open weights
Ungated unquantized bfloat16 open weights on Hugging Face at microsoft/Aion-1.0-Instruct-2.3B. Official model card: the unquantized model is published first, to be followed by quantized versions (Microsoft model card)
Parameters
2.3B
Type
Official model card: an instruction-tuned causal language model and foundational on-device small language model for text and operating-system use cases, from Microsoft's Windows Applied Sciences Group. Official overview: about 2.29B parameters with tied input/output embeddings, 28 layers, GQA (24 query heads, 8 key/value heads), a hybrid of 21 sliding-window layers (window 2048) and 7 full-attention layers, RoPE with YaRN scaling on full-attention layers, bfloat16 precision, and a knowledge cutoff of December 2025 (Microsoft model card)
Context
Official model card: 32,768 tokens (Microsoft model card)
Languages
Official model card: English, German, Spanish, Portuguese, Italian, French, Chinese, and Japanese (Microsoft model card)
Hosted access
Official Edge blog: generally available as Microsoft Edge's built-in small language model, replacing Phi-4-mini in Edge, with an origin trial for the Prompt API so sites can call it in Edge. The blog also describes a developer preview of an experimental Decisions API powered by Aion Instruct in Edge Canary and Dev. No official cloud API or API pricing is stated; the model is published for local use (Microsoft Edge blog and model card)
Windows plan
Official model card, as the lab's stated plan: a further optimized version of Aion Instruct, coupled with Microsoft's content moderation model, will become the Windows OS SLM by the end of January 2027. Official Edge blog: it is coming to Windows later this year. This Hugging Face release is not described as already replacing Phi Silica on Windows (Microsoft model card and Edge blog)
License
MIT License (Microsoft model card and LICENSE file)

Tags marked unsure are not stated on the lab's own pages.

Next steps

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

  • Publicly announced with a source
  • Usable by the public (API or weights)
  • Listed on an independent board we track
  • Enough results for a rank in a category
  • Holds across two snapshots (crowns only)
Self-reported

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: self-reported scores of 55.15 on MMLU, 33.35 on MMLU-Pro, 45.51 on MMMLU, 28.18 on GPQA, 74.91 on GSM8K (loose matching), 49.56 on GSMPlus, 48.00 on MATH-500, 83.90 on IFEval (mean of prompt-level strict and instruction-level strict), 32.00 on IFBench (prompt-level loose), 64.70 on Multi-IF, and 43.18 chrF++ on translation (official model card; lab-run with EleutherAI lm-evaluation-harness or each benchmark's official implementation).

Receipts

Sources

LABHybrid intelligence on the Web: Introducing Aion Instruct and Prompt API in Microsoft Edgeblogs.windows.com/msedgedev/2026/11/19/hybrid-intelligence-on-the-web-introducing-aion-instruct-and-prompt-api-in-microsoft-edge/ · Nov 19, 2026LABAion 1.0 Instruct (Microsoft model card)huggingface.co/microsoft/Aion-1.0-Instruct-2.3B

Tracking continues until an independent board lists it.