Primary use stated or inferred cautiously from official documentation.
Aion 1.0 Instruct 2.3B
Microsoft Windows Applied Sciences Group text model published on November 18. It has about 2.29 billion parameters and MIT-licensed BF16 weights. Designed for summarization, rewriting, chat and document understanding in eight languages, including Inferama’s six languages.
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
Limits and integration
| API ID | microsoft/Aion-1.0-Instruct-2.3B |
|---|---|
| Model type | Text and reasoning |
| Access model | Open weights |
| License | MIT |
| Deployment | Self-hosted |
| Release | 18 NOV 2026 |
| Knowledge cutoff | December 2025 |
| Entrance | Text |
| Exit | Text |
| Context window | 32,768 tokens |
| maximum output | Text; separate output maximum unpublished |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · Transformers · ONNX Runtime · WebGPU |
Local text-processing prototypes on compatible computers and browsers. The release adds an ONNX/WebGPU format with INT4 weights and FP16 activations and cache, around 1.3 GB. It is a deployment of the same model, not a duplicate profile.
The public version is not an already shipped Windows feature. It has the knowledge, reasoning and safety limitations of a small model. Transformers needs custom code: review it before loading. Local privacy also depends on the application, its telemetry and its connections.
What it can do
Orientation
Local text-processing prototypes on compatible computers and browsers. The release adds an ONNX/WebGPU format with INT4 weights and FP16 activations and cache, around 1.3 GB. It is a deployment of the same model, not a duplicate profile.
Context
32,768 tokens · Text; separate output maximum unpublished
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
Weights and the ONNX format are published under MIT. No official API price is verified. Local execution consumes memory, energy and compute; file size is not total required memory. Quantization may change accuracy and performance.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
How to read the results
Microsoft reports 55.15% MMLU and 83.90% IFEval. It uses lm-evaluation-harness when an implementation exists and the benchmark’s official implementation otherwise. IFEval averages prompt-level and instruction-level strict scores. These figures do not automatically apply to INT4 or every language.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| MMLU | 55.15% | Published score | View source ↗ |
| IFEval | 83.90% | Mean of prompt-level and instruction-level strict accuracy | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
Chronology
microsoft/Aion-1.0-Instruct-2.3B
BF16 checkpoint with custom code and MIT license.
microsoft/Aion-1.0-Instruct-2.3B-ONNX-WebGPU
ONNX INT4 format, revision 1677687f; does not add a model to the count.