Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license
Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.
Aion 1.0 Instruct 2.3B at a glance
How Aion 1.0 Instruct 2.3B ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| GPQA | 28.2% | 162 of 164 | Hugging Face model card, Nov 2026 ↗ |
| MMLU-Pro | 33.4% | 76 of 76 | Hugging Face model card, Nov 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
Microsoft has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model built for on-device use in Windows, on Hugging Face under the MIT license. The model supports a 32,768-token context window and has a knowledge cutoff of December 2025.
Microsoft's Windows & Devices Applied Sciences Group describes Aion Instruct as the foundational small language model (SLM) in the Windows ecosystem, optimized for on-device AI, text and operating system use cases. According to Microsoft, a further optimized version, paired with Microsoft's content moderation model, will become the Windows OS SLM by the end of January 2027. The Hugging Face release is meant as a preview of that model and a way to collect user feedback. The unquantized model is published first, with quantized versions to follow.
Specifications
- Parameters: about 2.29B (tied input/output embeddings)
- Architecture: causal language model, 28 layers, hidden size 2048, MLP intermediate size 8192 (SiLU)
- Attention: grouped-query attention with 24 query heads, 8 key/value heads, head dimension 128
- Attention pattern: hybrid, with 21 sliding-window layers (window 2048) and 7 full-attention layers
- Positional encoding: RoPE (θ = 32000), with YaRN scaling (factor 16) on full-attention layers
- Context length: 32,768 tokens
- Vocabulary: 200,029 tokens
- Precision: bfloat16
- Languages: English, German, Spanish, Portuguese, Italian, French, Chinese, Japanese
- Library formats listed: PyTorch, ONNX, GGUF
- Pricing: not applicable to the weights, which are MIT-licensed
The model uses custom modeling code, so trust_remote_code=True is required. It was exported with transformers 5.x. The chat template ends every turn with <|end|>, which is also the EOS token, and supports tool calling through apply_chat_template(..., tools=tools).
Training data
Microsoft says the training data extends the data used for Phi Silica, the previous generation Windows OS SLM. It combines rigorously quality-filtered public documents with newly created synthetic data covering math, coding, common sense reasoning and general world knowledge. Multilingual data makes up about 6% of the total.
Benchmarks (reported by Microsoft)
Microsoft ran benchmarks with EleutherAI's lm-evaluation-harness where a task implementation existed, and otherwise used each benchmark's official implementation. GPT-4o served as judge for LLM-judged benchmarks.
| Benchmark | Score |
|---|---|
| MMLU | 55.15 |
| MMLU-Pro | 33.35 |
| MMMLU | 45.51 |
| GPQA | 28.18 |
| GSM8K (loose matching) | 74.91 |
| GSMPlus | 49.56 |
| MATH-500 | 48.00 |
| IFEval | 83.90 |
| IFBench (prompt-level loose) | 32.00 |
| Multi-IF | 64.70 |
| Translation (chrF++) | 43.18 |
The IFEval score is the mean of prompt-level strict and instruction-level strict results.
Stated limitations
The model card says Aion Instruct was not trained to excel at world knowledge and can miss fine details, produce misstatements, or fail at complex multi-step reasoning. It may be susceptible to adversarial attacks. Microsoft names XPIA, DIC and ECI as key safety metrics to improve in future releases, and says output quality may differ by language. It lists out-of-scope uses including generating harmful or misleading content and automated decision-making in critical contexts such as medical diagnoses or legal judgments.
What this means
Aion Instruct shows the model Microsoft plans to ship as the Windows OS SLM, released openly ahead of general availability in Windows by the end of January 2027. The MIT license and the ONNX and GGUF formats allow developers to test it locally now. The scores show a split profile: IFEval (83.90) and GSM8K (74.91) are comparatively strong, while GPQA (28.18), MMLU-Pro (33.35) and IFBench (32.00) are low, consistent with Microsoft's own guidance that the model targets summarization, rewriting, chat and document comprehension rather than knowledge-heavy or complex reasoning tasks. Microsoft's plan to pair the final Windows model with a separate content moderation model reflects its acknowledgment that a 2.3B model cannot handle all safety nuances alone.
Related Articles
Bria releases RMBG-3.0, a 4-step background remover built on Fibo 2.0, with open weights for non-commercial use
Bria AI has released RMBG-3.0, the third generation of its background removal models and the first built on a generative foundation. It is Fibo 2.0, fine-tuned and distilled to produce an alpha matte in 4 steps. Bria claims it ranks first among six systems in its own benchmark of 234 hard images, with an average rank of 1.51.
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
Apple releases FastFrames-L, a 14B two-step video model distilled from Wan2.1-14B for interactive 480p generation
Apple has released FastFrames-L (14B) and FastFrames-S (1.3B), autoregressive two-step video diffusion models that generate 480p video at interactive latency. FastFrames-L is distilled from Wan2.1-14B, and the repository lists 35.4 ms total latency for FastFrames-S on one GPU.
Reson8 releases 500M-parameter Rumble-2-Flash speech model as free Apache 2.0 open weights
Amsterdam-based Reson8 (Resonate Labs B.V.) has released Rumble-2-Flash, a 500 million-parameter transcription model covering more than 40 languages, as a free Apache 2.0 open-weights download. The company claims a 7.09% average word error rate across seven languages on Hugging Face's Open ASR Leaderboard multilingual tests.
Comments
Loading...