LATEST
AI Tools

Microsoft Aion Instruct Is Live: A Free 2.3B Model Inside Edge

Updated November 20, 2026
17 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of a laptop screen showing programming code against a dark blue-lit background

Quick Answer & Key Takeaways

  • • Microsoft made Aion Instruct generally available on November 19, 2026, as a free 2.3B-parameter open-weights model under the MIT license.
  • • It replaces Phi-4-mini as the built-in small language model in Microsoft Edge, and developers can call it from live websites through a new Prompt API origin trial.
  • • The weights cost nothing to download, the model runs on-device with no per-token bill, and a further optimized version becomes the Windows OS small language model by the end of January 2027.

Your web browser now ships with an AI model inside it. On November 19, 2026, Microsoft made Aion Instruct generally available as a free, open-weights small language model, and it replaces Phi-4-mini as the built-in model in Microsoft Edge. Developers can now call that on-device model from live sites through a new Prompt API origin trial, with no cloud round trip and no per-token bill. This article covers what Aion Instruct is, what Microsoft's own benchmarks say about it, what it costs to run, how to try it yourself, and where it falls short.

What Aion Instruct is

Aion Instruct is a 2.3 billion parameter instruction-tuned language model from Microsoft's Windows Applied Sciences Group. Microsoft describes it as the foundational small language model of the Windows ecosystem, built for on-device text work: summarization, rewriting, document question answering, and chat. It ships under the MIT license, which means anyone can download the weights, use them commercially, and build on them without paying Microsoft anything.

The November 19 announcement carries three parts, and they have different maturity levels, so keep them separate. Aion Instruct itself is generally available, both as open weights on Hugging Face and as Edge's built-in small language model. The Prompt API, which lets websites call that built-in model, is in an origin trial, so developers can test it on live sites but it is not a finished standard. The Decisions API, a separate experimental interface for structured answers, is a developer preview in Edge Canary and Dev only. Microsoft says both the API and its implementation may change.

In Edge, Aion replaces Phi-4-mini, the 3.8 billion parameter model that held the built-in slot before it. Microsoft says Aion is smaller, faster, and available on many more devices than Phi, while delivering comparable quality on common web tasks such as text synthesis, summarization, instruction following, and structured outputs. That comparison comes with a gap worth naming: Microsoft published no head-to-head benchmark numbers for Aion against Phi-4-mini, so "comparable quality" is the company's claim, not a measured result. The two models' own cards do publish scores, but they were run on different evaluation setups, which we cover below.

Edge downloads Aion only when a site or feature first needs it, and only on devices that can run it well. Users stay in control: turning off on-device AI in Edge settings removes the model and stops future downloads. Microsoft also says Aion is coming to Windows later this year, and the model card adds the detail that a further optimized version of Aion, paired with Microsoft's content moderation model, will become the Windows OS small language model by the end of January 2027. The Hugging Face release is the unquantized model; quantized versions will follow.

What it costs: nothing, with one honest caveat

The price of Aion Instruct is zero. The weights are free to download from Hugging Face, the MIT license carries no fee or royalty, and running the model on your own hardware generates no per-token bill. There is no API key to buy, no free tier to exhaust, and no subscription you must pay for to use it. That is the whole pricing story, and it is the model's main selling point.

The honest caveat is that free weights are not free computing. The single BF16 weights file is about 4.6 GB, according to a read of the repository file listing, so the download itself is the first cost. Then you need hardware that can hold and run a 2.3 billion parameter model, plus the electricity to run it. A third-party guide puts the requirement at 8 GB or more of VRAM for the full model, and Microsoft says quantized versions are coming, which will cut both the download and the hardware bar. On a laptop you already own, though, the marginal cost of a summary or a rewrite is effectively nothing.

To see what that saves, compare one million input tokens plus one million output tokens, the standard unit for small-model work, against the hosted small models a developer would otherwise call:

ModelInput per 1M tokensOutput per 1M tokens1M in + 1M out
Aion Instruct (local)$0.00$0.00$0.00
Muse Ember (Meta Model API, up to 200K input)$0.10$0.50$0.60 (about £0.46, about C$0.86)
Claude Haiku 5.5$0.10$0.50$0.60 (about £0.46, about C$0.86)
GPT-6 Luna$0.10$0.50$0.60 (about £0.46, about C$0.86)
Gemini 3.6 Flash$0.75$3.75$4.50 (about £3.43, about C$6.45)

Currency conversions use November 20, 2026 market rates of 1 USD = 0.7628 GBP and 1 USD = 1.4323 CAD, and are approximate. The hosted rates are list prices from each vendor: Muse Ember's rates come from Meta's own announcement, and Gemini 3.6 Flash's rates come from Google's pricing page, where the $0.75 and $3.75 rates run through December 31, 2026 and then double to $1.50 and $7.50. Two details change the comparison. First, Muse Ember's rates triple above 200,000 input tokens, to $0.30 input and $1.50 output per million, and that higher rate applies to the whole request. Second, Aion's zero only holds if you already own the hardware. If you are buying a machine to run it, that purchase is the real price, and a hosted model at $0.60 per million tokens is cheaper until your volume grows large. At 10 million tokens a day, though, the hosted small models cost $6 to $45 a day while Aion costs nothing in tokens, which is the point of running a model locally. We covered Meta Muse Ember at launch, including its 200,000-token price step.

What Microsoft's own benchmarks say

Microsoft published a full evaluation table on the Aion model card. Every score below is Microsoft's own run, mostly using EleutherAI's lm-evaluation-harness, with GPT-4o as the judge where a judge was needed. No independent lab has reproduced these numbers yet, and community uptake is still small: the Hugging Face repository had passed 1,500 downloads by November 20, two days after the weights appeared. Read the scores as Microsoft's account of its own model, not as verified results.

CategoryBenchmarkScore
KnowledgeMMLU55.15
KnowledgeMMLU-Pro33.35
KnowledgeMMMLU45.51
KnowledgeGPQA28.18
MathGSM8K74.91
MathGSMPlus49.56
MathMATH-50048.00
Instruction followingIFEval83.90
Instruction followingIFBench32.00
Instruction followingMulti-IF64.70
TranslationchrF++43.18

The pattern is clear, and Microsoft does not hide it. Aion is strongest where a small model earns its keep: following formatting instructions (IFEval 83.90) and grade-school math word problems (GSM8K 74.91). Those are the skills a browser model needs for rewriting a paragraph, summarizing a page, or following a strict output format. Knowledge is the weak spot. MMLU-Pro at 33.35 and GPQA at 28.18 are low scores, and the model card says so directly: the model was not trained to excel at world knowledge, and small models lack the storage and reasoning depth to remember facts accurately. If you ask Aion a hard science question, expect a confident-sounding answer that may be wrong.

The safety section is unusually candid for a model card. Microsoft says Aion may be susceptible to adversarial attacks, and it names three metrics as the key ones to improve in future releases: XPIA, which measures prompt-injection resistance, plus DIC and ECI. In plain terms, the model can be talked into doing things it should not do, and Microsoft knows it. The card also bars automated decision-making in critical contexts such as medical diagnoses or legal judgments. A local content moderation model will sit alongside Aion when it becomes the Windows OS model in January 2027, which reads as Microsoft's answer to exactly this weakness.

Aion vs Phi-4-mini: what actually changes

Phi-4-mini is the model Aion displaces in Edge, so the comparison matters more than any comparison with a large cloud model. Here is what each model's own card states:

SpecAion InstructPhi-4-mini-instruct
ParametersAbout 2.29B3.8B
Context window32,768 tokens128,000 tokens
LicenseMITMIT
Languages8 (English, German, Spanish, Portuguese, Italian, French, Chinese, Japanese)Multilingual, English-first
Knowledge cutoffDecember 2025Not stated on card
MMLU55.1567.3
MMLU-Pro33.3552.8
GSM8K74.9188.6
Role in EdgeBuilt-in model from November 2026Previous built-in model

Two warnings before you read that table as a verdict. First, the two cards used different evaluation setups: Phi-4-mini's MMLU was 5-shot and its GSM8K was 8-shot with chain of thought, while Aion's card does not state matching shot counts, so the gap is directional, not a controlled head-to-head. Second, the trade is not one-directional. Aion is about 40 percent smaller, which is why Microsoft says it runs on many more devices, but it gives up three quarters of Phi-4-mini's context window: 32,768 tokens against 128,000. A 32K window still holds multi-page documents and long email threads, according to the model card, but a 128K window holds far more. Aion also scores lower on every knowledge and math benchmark Microsoft published for both models, even allowing for the different harnesses. The honest summary is that Microsoft traded capability for reach: a smaller model that fits on more machines, with less knowledge and a shorter memory, in exchange for being available wherever Edge runs.

The Prompt API origin trial: AI on your live site

The Prompt API is the reason Aion matters to web developers rather than only to model collectors. It is a browser API that lets a website call the built-in small language model directly, so a page can offer writing help, summarization, or rewriting without sending the user's text to a server. Until now it was a developer preview available only in Edge Canary and Dev. With the origin trial announced November 19, developers can register, receive a trial token, and call the API from a live site in Microsoft Edge to test it with a real audience.

Microsoft frames the design as hybrid intelligence. A site uses the on-device model when it is available and well-suited, and switches to a configured cloud provider when the local model is missing or the task needs more power. That fallback matters because Edge downloads Aion only on devices that can run it well, so a share of visitors will not have the model at all. The trial is also a standards play: Microsoft says feedback on the API's design, the model's performance, and interoperability problems across browsers and models will inform discussions with other browser vendors and the web community as the Prompt API is considered for standardization. To try it, register for the Prompt API origin trial, read the developer documentation, and test the playground sample. The trial carries no published price, and Microsoft describes it as a way to evaluate features under real-world conditions, not as a paid product.

One related development is worth knowing if you build agents with large tool libraries. Google shipped a feature this week that defers tool schemas until an agent needs them, cutting the token cost of big tool sets, and the same logic applies here: a browser model with a 32K window cannot afford to carry hundreds of tool definitions. We covered Google's tool retrieval feature in detail.

The Decisions API preview: answers without text

The third announcement is the most experimental and, for some builders, the most interesting. The Decisions API, powered by Aion Instruct, answers yes or no, multiple-choice, or scoring questions about a piece of text and returns probabilities and a confidence value in a single pass, without generating any text. Microsoft's example asks three questions about a product review at once: what the review is mainly about, how satisfied the customer is on a 1 to 5 scale, and whether the customer is asking for help.

// Ask three questions about a product review in a single pass
const model = await DecisionModel.create({
  "context": "Customer reviews for an online electronics store",
  "questions": [
    {
      "id": "topic",
      "type": "choice",
      "prompt": "What is the review mainly about?",
      "options": [
        { "label": "quality", "description": "How the product works or feels" },
        { "label": "shipping", "description": "Delivery, packaging, or delays" },
        { "label": "price", "description": "Cost or value for money" }
      ]
    },
    {
      "id": "rating",
      "type": "score",
      "prompt": "How satisfied is the customer, from 1 to 5?"
    },
    {
      "id": "needs_reply",
      "type": "boolean",
      "prompt": "Is the customer asking for help?"
    }
  ]
});

// Get a label, probabilities, and a confidence value for each question
const result = await model.decide(
  "Love the headphones, but the box arrived crushed. Can I get a replacement?"
);

The use cases Microsoft names are understanding user intent, sorting and routing user-generated content, and scoring and ranking items in real time, all with a few lines of client-side JavaScript. The important limits are that the Decisions API is available only in Edge Canary and Dev, it is a developer preview, and Microsoft says to expect both the API and its implementation to change as it learns from developer feedback. It is also a single-purpose tool: it classifies and scores text, it does not write, summarize, or chat. Google's own Decisions API takes a similar approach in the cloud, and we covered it at launch: Gemini Decisions API: 1,000 Questions for $0.10, No Output Charge. Microsoft's version runs on-device, which is the difference that matters for latency and privacy.

How to run the weights yourself

You do not need Edge to use Aion. The weights are on Hugging Face at microsoft/Aion-1.0-Instruct-2.3B, and the model card includes a quickstart that runs with the Transformers library. Two details in that quickstart deserve attention: the model uses custom modeling code, so trust_remote_code must be set, and it was exported with Transformers 5.x, so an older Transformers install may not load it.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "microsoft/Aion-1.0-Instruct-2.3B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "user", "content": "Give me a short introduction to large language models."},
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

The card lists PyTorch, ONNX, and GGUF as supported libraries, and the model supports tool calling through its chat template, with tool schemas injected into the system prompt and calls emitted in a tagged format. A community GGUF build already exists, but it is early: one third-party test on November 20 found that llama.cpp could not yet load the model because the architecture is not supported, so do not plan a llama.cpp workflow around Aion until that support lands. Quantized versions from Microsoft are promised but have no date. For a sense of what local models can already do on a laptop, we covered Recursive Flint, a free 40B coding model built to run on a laptop.

Before you build on it: the catches

Aion is a real release with real limits, and most of them are printed on its own model card. The first is that no independent benchmark exists yet. Every number in this article comes from Microsoft's own evaluation runs, and the repository had only about 1,500 downloads two days after release, so there is no community verification to lean on. Treat the scores as a starting point and test the model on your own tasks before trusting it.

The second catch is knowledge. Aion was not trained to excel at world knowledge, and its MMLU-Pro and GPQA scores show it. Use it to rewrite, summarize, and follow formats, not as a fact engine. The third is the context window. At 32,768 tokens it handles long documents, but it is a quarter of what Phi-4-mini offered, so long-context work that fit in the old built-in model may not fit in the new one.

The fourth catch is safety. Microsoft's own card says the model may be susceptible to adversarial attacks and names prompt-injection resistance as a metric to improve. Do not put Aion in front of untrusted text, such as web pages or user-submitted content, without guardrails, and do not use it for decisions that affect health, money, or legal status. The fifth is maturity. The Prompt API is an origin trial and the Decisions API is a Canary and Dev experiment, so code written against either one may need changes. The sixth is scope: Aion is text only. It does not see images, hear audio, or generate video, and it only runs where Edge or your own hardware runs it. Chrome and Safari do not ship this model, and the Windows version arriving in January 2027 is a different, further optimized build with its own moderation model, not the file on Hugging Face today.

One more honest note: this article is a research review based on Microsoft's announcement, the model card, and independent coverage. We did not run the model ourselves, so nothing here claims hands-on testing. The numbers, dates, and limits above all come from the sources listed at the end.

Who should use it now, and who should wait

Use Aion now if you build for Edge or Windows and want free on-device text features. Summarization, rewriting, and structured outputs are exactly what the model is tuned for, and the Prompt API origin trial is the way to test those features on a live site without paying per token. Use it now if privacy matters to your users, since text processed on-device never leaves the machine. And use it now if you are a hobbyist with a capable GPU who wants a permissively licensed small model to experiment with, since the MIT license imposes no conditions beyond keeping the notice.

Wait if you need knowledge-heavy answers, long-context work beyond 32K tokens, or a stable production API. The Prompt API is a trial, the Decisions API is an experiment, and the model that lands in Windows in January 2027 will differ from the weights published today. Wait too if your users are not on Edge, since the browser integration is the whole point of this release, and the weights alone are one small model among many. For teams already wiring large tool libraries into agents, the more urgent question is token cost, and Google's data tooling for agents is a useful comparison: we covered Google's Data Agent Kit and its free MCP tools earlier.

Frequently asked questions

Is Microsoft Aion Instruct really free?

Yes. The weights are published under the MIT license on Hugging Face, free to download and free to use commercially, with no per-token charge when you run the model on your own hardware. The real costs are the roughly 4.6 GB download, hardware that can run a 2.3 billion parameter model, and electricity.

How do I use Aion Instruct in Microsoft Edge?

You do not need to install anything. Edge downloads Aion automatically the first time a site or feature needs it, and only on devices that can run it well. You can turn off on-device AI in Edge settings, which removes the model and stops future downloads.

Can I run Aion Instruct on my own computer?

Yes. The weights are at microsoft/Aion-1.0-Instruct-2.3B on Hugging Face, and the model card gives a Transformers quickstart. You need about 4.6 GB for the BF16 weights and a GPU with 8 GB or more of VRAM according to a third-party guide, plus trust_remote_code set because the model uses custom code. Quantized versions from Microsoft are coming but have no date.

Is Aion better than Phi-4-mini?

It is smaller and, according to Microsoft, faster and available on more devices, but it scores lower on every knowledge and math benchmark Microsoft published for both models, and its 32K context window is a quarter of Phi-4-mini's 128K. Microsoft published no head-to-head benchmark, so the fair reading is a trade: more reach, less capability.

What is the Prompt API origin trial?

It is a registration-based trial that lets developers call Edge's built-in Aion model from a live website, instead of only in Canary and Dev builds. You register, receive a token, and test on-device AI features with real users. Microsoft publishes no price for the trial, and feedback from it will inform web standardization discussions.

When does Aion come to Windows?

Microsoft's Edge blog says Aion is coming to Windows later this year. The model card is more specific: a further optimized version of Aion, paired with a Microsoft content moderation model, will become the Windows OS small language model by the end of January 2027.

Aion Instruct is Microsoft's boldest free AI bet: a capable small model, free under a permissive license, running inside a browser millions of people already have. Its limits are printed on its own model card, and they are real: weak knowledge, a 32K window, and prompt-injection resistance still to improve. If you build for Edge, try the Prompt API playground first, read the model card before you trust the benchmarks, and test the model on your own tasks before you build on it.

Sources

17 min read
26 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more