Aion 1.0 Instruct 2.3B
AvailableSmall on-device language model from Microsoft's Windows Applied Sciences Group, published on Hugging Face Nov 18 2026 under the MIT License. About 2.29B parameters (dense, 28 layers, hybrid sliding-window and full attention, GQA), 32,768-token context, December 2025 knowledge cutoff, eight languages, tool calling; training data extends that of Phi Silica. Microsoft says a further optimized version, paired with a content moderation model, will become the Windows OS small language model by the end of January 2027. BF16 weights plus an INT4 ONNX WebGPU build; custom code (trust_remote_code). Model card reports MMLU 55.15 and IFEval 83.90 (vendor-reported).
Specifications
- License
- Open source · MIT
- Weights
- Downloadable
- Architecture
- Dense
- Parameters
- 2.29B
- Context window
- 33K tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- —
- Modalities
- Text
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.