Tasks and inputs
Aion 1.0 Instruct 2.3B is listed for Question answering, Summarization, Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 19, 2026
Aion 1.0 Instruct 2.3B is Microsoft's MIT-licensed on-device small language model and Phi Silica successor for Windows, with a 32K context.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights under MIT; no hosted API price.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Phi Silica has a successor. Microsoft's Windows Applied Sciences Group published Aion 1.0 Instruct 2.3B on Hugging Face on 18 November 2026, describing it as the foundational small language model for the Windows ecosystem, and said that a further-optimised build, paired with a local content moderation model, will ship as the Windows OS model by the end of January 2027.
So this is a preview of something you will eventually call through Windows itself. The weights, though, are already open under MIT.
| Detail | Value |
|---|---|
| Repository | microsoft/Aion-1.0-Instruct-2.3B |
| Parameters | ~2.29B, tied input/output embeddings |
| Layers / hidden size | 28 / 2048 |
| Attention | GQA, 24 query and 8 KV heads; 21 sliding-window layers (2,048 window) plus 7 full-attention layers |
| Context | 32,768 tokens |
| Vocabulary | 200,029 |
| Precision | bfloat16 |
| Knowledge cutoff | December 2025 |
| Languages | English, German, Spanish, Portuguese, Italian, French, Chinese, Japanese |
| License | MIT |
An int4 ONNX build for WebGPU appeared the same day, and Microsoft says further quantised versions will follow the unquantised release.
Loading needs trust_remote_code=True, because the architecture ships as custom modelling code, and the checkpoint was exported with transformers 5.x. Turns end with <|end|>, which doubles as the EOS token. Tool calling works by injecting schemas between <|tool|> tags; the model answers with JSON wrapped in <|tool_call|>, and results go back in a user turn prefixed <|tool_response|>.
For the Windows path there is a separate developer validation package. It targets Copilot+ PCs only: Snapdragon ARM64 through QNN, Intel Lunar Lake through OpenVINO, and AMD Ryzen AI through VitisAI. No NPU provider, no fallback; it will not drop to the CPU. That package is for validation, not redistribution, and its output carries invisible watermarking that can surface when pasted into some applications.
Microsoft frames the model around document comprehension, summarisation, rewriting, question answering and chat on the device, with data never leaving the machine. Thirty-two thousand tokens is enough for a long email thread or a multi-page spec. It is not enough for a codebase.
Its publisher-reported scores set expectations honestly: 55.15 on MMLU, 33.35 on MMLU-Pro, 28.18 on GPQA, 74.91 on GSM8K and 83.90 on IFEval. Those numbers say "follows instructions well, knows a moderate amount, reasons shallowly", and Microsoft says as much, warning that it was not trained to excel at world knowledge and may miss detail on complex multi-step reasoning. Multilingual material made up roughly 6% of training data, so expect weaker output outside English.
If you are comparing it with an earlier Microsoft small model such as Phi-4-mini, note that this is a different architecture rather than a refresh. Existing Phi Silica LoRA adapters will not load. New ones have to be trained with Foundry Toolkit.
Pass on most rewriting and summary prompts but fail on facts? Then it suits on-device drafting, with a larger model kept for anything that must be correct.
More from the lab is collected on the Microsoft provider page.
Compare the model on the work you need to complete.
Aion 1.0 Instruct 2.3B is listed for Question answering, Summarization, Text generation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 32.77K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under MIT. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Gemini 4 Neon is Google's mid-tier Gemini 4 model for complex coding and analysis, sitting between Hydrogen and Carbon at $3/$15 per 1M tokens.
OpenAI
GPT-6.2 Astra is OpenAI's November 2026 flagship for hard professional, scientific and agentic work, at $8/$40 per million tokens with a 1.05M context.
OpenAI
GPT-6.2 Sol is OpenAI's everyday model for coding, agents and writing: $2/$10 per million tokens, 1.05M context, released 17 November 2026.
Tencent
Tencent Hy4 Flash is the cost-focused companion to Hy4, keeping vision and agent skills at a far lower token price. Availability is still unsettled.
Alibaba
Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.
Sarvam
Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.