Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

Open-Weights Releases

Microsoft Aion 1.0 Instruct: MIT open weights and a Prompt API in Edge

Microsoft published MIT open weights for Aion 1.0 Instruct, a 2.3B model, and made it Edge's built-in small language model with a Prompt API trial.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

A slim laptop on a plain desk with a dim screen glow, a low server cabinet with closed doors behind it — the browser-served 2.3B-parameter open-weights model running where the user sits. Illustration: DeAI
A slim laptop on a plain desk with a dim screen glow, a low server cabinet with closed doors behind it — the browser-served 2.3B-parameter open-weights model running where the user sits. Illustration: DeAI

Microsoft has published open weights for Aion 1.0 Instruct under the MIT license and made the 2.3B-parameter model the built-in small language model of Microsoft Edge, with a Prompt API origin trial for live sites. That puts a permissively licensed open-weights model in front of one of the largest desktop browser footprints on earth.

Key facts

  • The Hugging Face model card specifies ~2.29B parameters, a 32,768-token context window, an eight-language set (English, German, Spanish, Portuguese, Italian, French, Chinese, Japanese), and the MIT license.
  • Per the same card, Microsoft's self-run benchmarks put MMLU at 55.15, GSM8K at 74.91, IFEval at 83.90, and GPQA at 28.18, run with EleutherAI's lm-evaluation-harness and GPT-4o as judge where a benchmark needed one.
  • Microsoft's Edge blog confirms Aion Instruct replaces Phi-4-mini as Edge's built-in small language model and opens a Prompt API origin trial so live sites can call the API from production code.
  • The experimental Decisions API enters developer preview: it answers yes/no, multiple-choice, and scoring questions about text in a single pass and returns probabilities with a confidence value, without generating text.
  • Microsoft says an optimized version paired with its content moderation model becomes the Windows operating system's built-in small language model by the end of January 2027.

What happened

The announcement came from the Edge team on November 19 in a post framed around "hybrid intelligence on the web." Three things shipped at once. Aion Instruct is generally available as an open-weights model on Hugging Face and as the browser's built-in small language model. The Prompt API, previously locked to Edge Canary and Dev channels, is now in an origin trial, which means a production website can register a token and use on-device inference with real users. And a new experimental Decisions API entered developer preview in the Canary and Dev channels.

The model itself is small and openly specified. The model card lists ~2.29B parameters with tied input/output embeddings, 28 layers, grouped-query attention with 24 query heads and 8 key/value heads, and a hybrid attention pattern of 21 sliding-window layers at a 2,048-token window plus 7 full-attention layers. Context runs to 32,768 tokens, the vocabulary to 200,029, and precision is bfloat16. Training data extends what Microsoft used for Phi Silica, the previous Windows on-device model, with multilingual data making up about 6% of it. Files ship in PyTorch, ONNX, and GGUF form, and the custom modeling code means trust_remote_code=True is required to load it in Transformers.

The browser integration is the part with real distribution. Edge downloads Aion only when a site or feature first needs it, and only on devices that can run it well. Users can turn off on-device AI in settings, which removes the model and stops future downloads. Microsoft describes Aion as "smaller, faster, and available on many more devices than Phi" while delivering comparable quality for summarization, instruction following, and structured outputs. That is the vendor's characterization; the released benchmark table is the only public evidence either way, and it is self-run.

Why it matters

Distribution is the story. An MIT-licensed open-weights model is now the default inference path inside a mainstream desktop browser, and the same weights sit on Hugging Face, so what Edge serves and what a developer self-hosts can be the same artifact. For teams building hybrid local-plus-cloud inference, the on-device tier now has a mainstream default, and Microsoft frames the Prompt API explicitly as one leg of a hybrid design: use on-device models when they suit, switch to a configured cloud provider when they don't.

That design has a consequence worth naming. Tasks that stay on-device never touch a provider's logs, and the dividing line is now a per-call choice the site makes in JavaScript. The Decisions API pushes the same logic further: classification-style questions answered in one pass, no generated text, probabilities and a confidence value back. For high-volume routing, moderation, and intent-detection workloads, that is a materially cheaper shape than generating tokens.

One caution belongs next to the enthusiasm. An on-device gateway to a browser vendor's AI APIs is also a walled garden in the making, since Edge's API surface serves Microsoft's built-in copy on Microsoft's terms. What blunts that here is the license: the weights are MIT, so the model itself is not the lock-in, and any developer unhappy with the browser path can run the same 2.3B checkpoint themselves.

Background

This is the general-availability step of a staged rollout. Edge shipped on-device AI with Phi-4-mini as the built-in model in May 2025 alongside the first Prompt API developer preview, and Aion-1.0-Instruct entered developer preview in June 2026. The November 19 post is the point where the experimental model, the experimental API, and live-site access line up.

It also lands in a crowded field of open small models. Liquid AI's LFM2.5-24B targets local agents at a much larger parameter count, Conway's underdog on-device bet argues for the same local-first thesis on Apple silicon, and Meituan's LongCat release shows Chinese labs shipping open weights at scale. Aion's distinction is license and placement: MIT rather than a community license, and a default seat in Edge rather than a download. Where self-hosting meets a private endpoint, the same trade-off between local control and provider convenience keeps recurring; this release moves one more tier of inference onto the user's side of that line.

What's next

Three dates and decisions to watch. The origin trial determines whether the Prompt API gets real production usage, and Microsoft says the trial feedback informs standardization discussions with other browser vendors, tracked in the MSEdgeExplainers GitHub discussion. The Decisions API implementation is expected to change during preview. And the model card puts the Windows operating-system integration, with the content moderation model attached, at the end of January 2027. On the verification side: no independent reproduction of the benchmark table exists yet, so the numbers above stay vendor-reported until someone else runs them.

Questions

What did Microsoft release as open weights in Aion 1.0 Instruct?
The weights of Aion 1.0 Instruct, a 2.3B-parameter instruction-tuned language model, published on Hugging Face under the MIT license in PyTorch, ONNX, and GGUF form. The card specifies a 32,768-token context window, an eight-language set, and bfloat16 precision.
What is the Prompt API in Microsoft Edge?
A web API that lets a site prompt the small language model built into Microsoft Edge, now Aion Instruct, from client-side JavaScript. It moved from Edge Canary and Dev previews to an origin trial on November 19, 2026, so live sites can register a token and use it.
Are Microsoft's Aion Instruct benchmarks verified?
No. The model card's scores, including MMLU 55.15 and IFEval 83.90, are Microsoft's own evaluations run with EleutherAI's lm-evaluation-harness. No independent reproduction has been published yet, so treat them as vendor-reported figures.
What is the Decisions API in Edge?
An experimental API in developer preview, powered by Aion Instruct, that answers yes/no, multiple-choice, and scoring questions about text in a single pass and returns probabilities with a confidence value without generating text.
When does Aion Instruct become the Windows built-in model?
Microsoft says an optimized version paired with its content moderation model will become the Windows operating system's built-in small language model by the end of January 2027. That is a vendor statement about future plans, not a shipped fact.

Sources

  1. Hybrid intelligence on the Web: Introducing Aion Instruct and Prompt API in Microsoft Edge — Microsoft Edge Blog
  2. microsoft/Aion-1.0-Instruct-2.3B model card — Hugging Face
  3. Decisions API explainer — GitHub
  4. Prompt a built-in language model with the Prompt API — Microsoft Learn
  5. Quantized Aion Instruct community models — Hugging Face

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →