what is aion instruct? microsoft's on-device model, explained
aion instruct is microsoft's small language model for on-device ai, published as aion 1.0 instruct with 2.3 billion parameters. on november 19, 2026, microsoft made it generally available as microsoft edge's built-in small language model, replacing phi-4-mini, and as open weights on hugging face under the mit license.
download the ReMynd desktop app to search text you saw in a focused window, by app and date.
what did microsoft announce?
microsoft's edge blog post lists three updates:
- general availability of aion instruct as an open-weights model on hugging face, and as microsoft edge's built-in small language model.
- an origin trial for the prompt api, so a site can use the prompt api and aion instruct on its live pages.
- a developer preview of an experimental decisions api, which returns probabilities and a confidence value for yes or no, multiple choice or scoring questions about text, without generating text.
the edge team says aion is smaller, faster and available on many more devices than phi, with comparable quality for common web tasks such as summarization, instruction following and structured outputs. it follows a developer preview that microsoft announced on june 2, 2026, which described the phi model edge used before as a 4b parameter language model.
what are aion instruct's specs?
microsoft's model card on hugging face lists these details.
| spec | what microsoft lists |
|---|---|
| type | causal language model (instruction tuned) |
| parameters | ~2.29b (tied input/output embeddings) |
| context length | 32,768 tokens |
| layers | 28 |
| vocabulary size | 200,029 |
| precision | bfloat16 |
| knowledge cutoff | 2025-12 |
| languages | english, german, spanish, portuguese, italian, french, chinese, japanese |
| license | mit |
the card says the model was primarily trained for general purpose text scenarios, including document comprehension, content analysis, question answering and natural language understanding. it supports tool calling through its chat template, and running it with the transformers library needs trust_remote_code=True because it uses custom modeling code.
what benchmarks does microsoft report?
the model card reports these scores, run with eleutherai's lm evaluation harness where a task implementation was available. microsoft says scores are percentages unless stated otherwise.
| benchmark | category | score |
|---|---|---|
| mmlu | knowledge | 55.15% |
| mmlu-pro | knowledge | 33.35% |
| gpqa | knowledge | 28.18% |
| gsm8k | math | 74.91% (loose matching) |
| math-500 | math | 48.00% |
| ifeval | instruction following | 83.90% (mean of prompt-level and instruction-level strict) |
these are microsoft's own numbers, the card shows no other model next to them, and we did not run the tests ourselves.
how does aion instruct compare with phi silica?
on windows, microsoft plans for aion instruct to replace phi silica. microsoft's windows developer blog lists phi silica at 3.8b parameters with an input context length limit of 3.5k tokens, and says microsoft is expanding aion to supported gpus, including nvidia's rtx series and amd's rdna4 generation.
its timeline:
- october 2026: a standalone package at aka.ms/tryaion for developers to start testing.
- december 4: planned rollout to windows insider preview devices.
- january 2027: planned rollout to retail devices, completing the transition from phi silica.
the same post says most existing ai features will keep working, with the exception of custom loras, and that microsoft added invisible watermarking to aion to fulfill text provenance requirements in specific regions.
what does microsoft edge need to run it?
microsoft learn's prompt api page says the api uses aion instruct, built into the browser, and lists these limits for the developer preview.
| requirement | what microsoft lists |
|---|---|
| operating system | windows 10 or 11; macos 13.3 or later; 64-bit (x64) linux with vulkan gpu drivers and microsoft edge 159 or later |
| storage | at least 20 gb free on the volume with your edge profile; below 5 gb, the model is deleted |
| gpu | 3 gb of vram or more |
| network | the model isn't downloaded on a metered connection |
the edge blog adds that edge downloads aion instruct only when a site or feature first needs it, and only on devices that can run it well. microsoft learn says that starting with microsoft edge version 158, text generated by the prompt api includes an invisible watermark.
how do you turn it off?
microsoft's on-device ai support page says to select settings and more > settings, select copilot and ai, then turn the on-device ai toggle on or off. you can also enter edge://settings/ai in the address bar.
when you turn off on-device ai, microsoft says edge deletes the locally stored ai model files, won't download ai models while the setting is off, and features that rely on them may stop working or have limited functionality. on-device ai is turned off by default in microsoft edge beta, dev and canary.
what are aion instruct's limits?
the model card is plain about them. it says small language models such as aion instruct can miss fine details, provide misstatements or show inability to follow complex multi-step reasoning, and that the model was not trained to excel at world knowledge.
- it says aion instruct may be susceptible to adversarial attacks.
- multilingual data is about 6% of its training data, and output quality may differ depending on the language.
- microsoft is publishing the unquantized model first, with quantized versions to follow.
where does ReMynd fit?
ReMynd records your screen on a mac, runs ocr locally with apple vision, and keeps a searchable index on your computer. text comes from the focused window only, and video covers windows you have not excluded.
in our view, a small on-device model works with the text you give it, so it does not know what you saw last week. ReMynd keeps that history searchable: press control space, search in plain language, and jump to the exact moment, provided ReMynd was running then and the window was not excluded. on a mac it needs macos 14 or later on an apple silicon mac.
ReMynd also runs on windows 11 (x64 or arm64). for running models on your own machine, see best local llm for mac.
limits of this guide
- it follows microsoft's edge blog post, the hugging face model card, the windows developer blog and microsoft learn, as of november 2026. we did not run the model.
- microsoft's edge post says aion is coming to windows later this year, while the windows developer blog's timeline above puts the retail rollout later; we report both as written.
- the benchmark table is a selection; the model card lists more scores.
download the ReMynd desktop app and search what you saw in a focused window, by app and date.