Original title: microsoft/Aion-1.0-Instruct-2.3B
An instruction model uploaded by the Microsoft Windows & Devices Applied Sciences Group on 11 18, positioned as a foundational small model for the Windows ecosystem, succeeding the previous generation system small model Phi Silica. It has approximately 2.29B parameters (with shared input and output embeddings) and 28 layers, of which 21 layers use sliding window attention with a window size of 2048, and 7 layers use full attention (GQA, with 24 query heads and 8 KV heads). It supports a context length of 32,768 tokens, has a vocabulary size of 200,029, and has a knowledge cutoff of 2025 12. It supports 8 languages including English, German, Spanish, Portuguese, Italian, French, Chinese, and Japanese, focusing on document understanding, summarization, rewriting, question answering, and dialogue, with support for tool calling formats. Official statement: An optimized version incorporating Microsoft's content moderation model will become the Windows system small model by the end of 2027 1. This release is the unquantized version, with the quantized version to follow. MIT License.
Why it matters Officially, the optimized version incorporating the content moderation model will become the built-in Windows system small model by the end of January next year; this is an early public preview
Rising Score 47 · Likes 67 · Downloads 788
Reasons
- Official self-testing using lm-evaluation-harness: MMLU 55.15, MMLU-Pro 33.35, GPQA 28.18, GSM8K 74.91, MATH-500 48.00, IFEval 83.90, Multi-IF 64.70, IFBench 32.00; no comparisons with other models are included in the card
- Training data expanded from Phi Silica's data: strictly filtered public documents plus new synthetic data for mathematics, code, and commonsense reasoning; multilingual data accounts for approximately 6%
- Positioned for local processing where data does not leave the device: multi-page documents, long email threads, and technical specifications can be placed entirely within the 32K context
Usage
- Transformers 5.x: Load microsoft/Aion-1.0-Instruct-2.3B using AutoModelForCausalLM / AutoTokenizer; trust_remote_code=True is required (custom model code modeling_aion_instruct.py), with dtype="auto" and device_map="auto"
- For conversations, use the repository's chat_template.jinja, where each turn ends with <|end|>; optional template variables include model_identity, knowledge_cutoff, and current_date
- Tool calling: Use apply_chat_template(..., tools=tools) to inject tool descriptions into the system prompt; the model outputs <|tool_call|>{"name": ..., "arguments": {...}}<|/tool_call|>, and tool results are returned to the user turn starting with <|tool_response|>; sample code is available at GitHub microsoft/Aion-Instruct-Preview-Sample
Risks
- Currently only BF16 safetensors are available; ONNX and GGUF are listed in the card, but the quantized version has not yet been uploaded
- Officially stated that it is not strong in world knowledge and may make errors in complex multi-step reasoning; safety metrics XPIA, DIC, and ECI still need improvement in future versions, and it is susceptible to adversarial prompt induction; a content moderation model will be configured separately during system integration
- Evaluations only include self-reported scores, with no comparisons against models of similar size; the proportion of multilingual data is low, so output quality for non-English languages may be inconsistent. This entry has not been run