ModelsNews
Aion Instruct: open 2.3B model replaces Phi in Edge
Microsoft has released Aion Instruct, a 2.3-billion-parameter model under the MIT license, which replaces Phi-4-mini as Edge's built-in small language model. The Prompt API has entered an origin trial, so you can now test it on a live site.
· 3 min read

What launched?
The Microsoft Edge team announced three updates on November 19 (Microsoft Edge Blog):
- Aion Instruct is generally available: as open weights on Hugging Face and as Edge's built-in small language model (SLM).
- The Prompt API has entered an origin trial, so you can call it from a live site in Edge.
- An experimental Decisions API is available as a developer preview, returning structured answers without generating text.
In Edge, Aion replaces Phi-4-mini. Microsoft says it is smaller, faster and runs on many more devices, with comparable quality for common web tasks such as summarization, text generation, instruction following and structured output. Edge downloads it only when a site or feature first needs it, and only on devices that can run it well. Users can turn it off in settings, which also removes the model.
What we know about the model
According to the model card (Hugging Face):
- About 2.29B parameters, 28 layers, 32K-token context, bfloat16 weights, MIT license.
- Hybrid attention: 21 sliding-window layers (2,048-token window) and 7 full-attention layers, with GQA.
- Languages: English, German, Spanish, Portuguese, Italian, French, Chinese and Japanese. Multilingual data makes up about 6% of training data.
- Training data extends the data used for Phi Silica (the previous Windows SLM); knowledge cutoff is December 2025.
- Reported scores: MMLU 55.15, MMLU-Pro 33.35, GSM8K 74.91, MATH-500 48.00, IFEval 83.90.
- It supports tool calling and needs trust_remote_code=True. The unquantized model ships first, with quantized versions to follow.
The model card says a further optimized version, paired with Microsoft's content moderation model, will become the Windows OS SLM by the end of January 2027.
Prompt API and Decisions API
Until now, the Prompt API was a developer preview in Edge Canary and Dev only. With the origin trial, you can test it on a live site with real users. Microsoft will use the feedback in discussions with other browser vendors, since the Prompt API is being considered for standardization.
The Decisions API answers yes/no, multiple-choice or scoring questions about text with probabilities and a confidence value, in a single pass and without text generation. Think intent detection, routing user-generated content or real-time ranking with a few lines of client-side JavaScript. Google and OpenAI offer a similar idea in the cloud (Gemini Decisions API); Edge's runs on the device.
Why it matters
A model built into the browser means simple AI features (summaries, tagging, routing) need no API key and no cloud bill, and the data stays on the device. Microsoft pushes a hybrid approach: use the local model when it is available and good enough, and fall back to the cloud otherwise. If your site is not in one of the eight supported languages, test quality first.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About


