Microsoft's Windows Applied Sciences Group released Aion 1.0 Instruct on November 18, 2026. It is a 2.29B parameter dense text model built for on-device use in Windows, with a 32,768 token context window and support for eight languages. Weights are published on Hugging Face under the MIT license, and Microsoft says an optimized version will become the Windows OS small language model by the end of January 2027. It is trained for summarization, rewriting, document question answering and chat.
A workable 2.29B-parameter dense language model from microsoft. A pragmatic middle-ground choice when you need open weights without a flagship-sized footprint. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
No benchmark data available for this model yet.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
See how different quantization levels affect VRAM requirements and quality for this model.
| Format | VRAM Required | Quality | |
|---|---|---|---|
| Q2_K | 1.9 GB | Low | |
| Q4_K_MRecommended | 2.4 GB | Good | |
| Q5_K_M | 2.7 GB | Very Good | |
| Q6_K | 2.9 GB | Excellent | |
| Q8_0 | 3.5 GB | Near Perfect | |
| FP16 | 5.7 GB | Full |
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | Speed | VRAM |
|---|---|---|---|
| NVIDIA RTX A2000NVIDIA | AA | 95.5 tok/s | 2.4 GB |
| AMD Radeon RX 7600 8GBAMD | AA | 95.5 tok/s | 2.4 GB |
| NVIDIA GeForce RTX 4060NVIDIA | AA | 90.2 tok/s | 2.4 GB |
| NVIDIA GeForce RTX 5060 Ti 8GBNVIDIA | AA | 148.5 tok/s | 2.4 GB |
| AMD Radeon RX 7700 XTAMD | AA | 143.2 tok/s | 2.4 GB |
Energy cost on NVIDIA RTX A2000 (~95 tok/s, Q4_K_M) vs flagship API pricing.
| Source | Cost per 1M tokens |
|---|---|
Local (energy only)Aion 1.0 Instruct 2.3B on NVIDIA RTX A2000 · ~95 tok/s · 70W | $0.024 |
GPT-6.2 SolOpenAI · in $2.00 · out $10.00 | $4.40 |
Claude Fable 5.2Anthropic · in $8.00 · out $40.00 | $17.60 |
Gemini 4 NeonGoogle · in $3.00 · out $15.00 | $6.60 |
Grok 4.5xAI · in $2.00 · out $6.00 | $3.20 |
API prices blended at 70% input / 30% output.
Hardware amortisation not included. Run the full ROI calculator for payback math.
Cheapest current cloud rentals with at least 2 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA GeForce RTX 3060Vast.ai · Spot · 12 GB VRAM | $0.06 |
NVIDIA GeForce RTX 3060Vast.ai · On-Demand · 12 GB VRAM | $0.06 |
NVIDIA GeForce RTX 5070Vast.ai · Spot · 12 GB VRAM | $0.07 |
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.08 |
NVIDIA RTX A4000Vast.ai · Spot · 16 GB VRAM | $0.08 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.
Cumulative cost of buying the smallest device in our directory that runs this model at 4-bit, compared with paying flagship APIs for the same token volume.
Assumes 1 million tokens a day, a 70/30 input and output mix and $0.12 per kWh. The device only runs as long as the workload needs. Use the ROI calculator for your own numbers.
Microsoft's Windows Applied Sciences Group released Aion 1.0 Instruct 2.3B on November 18, 2026, positioning it as the foundation for the next generation of local operating system intelligence on Windows. Built as the direct architectural successor to Phi Silica, this dense 2.29B parameter model drops parameter overhead from Silica's 3.8B footprint while expanding the context window from 3.5K to a production-ready 32,768 tokens. Released openly on Hugging Face under the permissive MIT license, an optimized iteration paired with system content moderation is slated to become the default in-box Windows small language model (SLM) by late January 2027.
Unlike generalist research checkpoints that require heavy quantization surgery to fit edge devices, Aion 1.0 Instruct is designed from day one for local AI execution. It is tailored to run concurrently alongside desktop applications without monopolizing system memory or compute budgets. By shipping unquantized weights alongside support for ONNX and GGUF, Microsoft provides developers with a transparent preview of the exact language foundation destined for consumer Copilot+ PCs.
For engineers seeking a modern local AI model with 2.29B parameters in late 2026, Aion occupies a strategic sweet spot. It targets document comprehension, summarization, structured data extraction, and tool execution entirely offline. Its small memory footprint allows developers to embed local language intelligence directly into desktop software, background daemons, and autonomous local agent loops with minimal hardware overhead.
Aion 1.0 Instruct is a standard dense transformer model containing 2.29 billion parameters. It avoids the sparse activation complexity of Mixture of Experts (MoE) architectures, ensuring predictable memory bandwidth utilization and consistent token generation latencies across both integrated NPUs and discrete GPUs.
The architecture features 28 transformer layers with a hidden dimension of 2048, utilizing tied input and output embeddings to preserve parameter budget for representational depth. Key architectural attributes include:
The 32K context window represents a roughly nine-fold expansion over Phi Silica. At 2.29B parameters, maintaining a dense context of 32,768 tokens in full 16-bit precision would typically threaten to outgrow the model weights themselves. Thanks to GQA, the KV cache overhead remains manageable on consumer hardware, enabling dense multi-document analysis without requiring offloading to host system RAM.
Aion 1.0 Instruct is tuned primarily as an instruction-following and system-task model. Rather than targeting deep academic math proofs or massive code synthesis, its pre-training and alignment prioritize high-utility text manipulation, parsing, and conversational routing.
Due to its compact parameter count, Aion 1.0 Instruct 2.3B hardware requirements are remarkably modest compared to 7B or 8B class models. It runs smoothly on entry-level hardware, integrated graphics, and standalone NPUs.
To run Aion 1.0 Instruct 2.3B locally, base weight footprints across common quantization formats scale as follows:
When operating at the full 32,768-token context boundary, the GQA-enabled KV cache adds approximately 1.0 GB to 1.5 GB of memory overhead depending on runtime cache precision (FP16 vs. Q8/Q4 cache).
The best quantization for Aion 1.0 Instruct 2.3B in general daily workloads is Q4_K_M. Because the model has 2.29B parameters, dropping down to sub-4-bit quantizations (such as Q3 or Q2) leads to noticeable degradation in instruction adherence and function-calling syntax. For mission-critical structured output and syntax-sensitive tasks, step up to Q8_0 or FP16, both of which still easily fit inside a standard 6 GB or 8 GB consumer GPU.
Aion 1.0 Instruct 2.3B tokens per second metrics vary by compute engine:
The quickest path to test the model locally is via Ollama or standalone llama.cpp. Once the GGUF weights are converted or pulled, running the model in Ollama requires only creating a standard Modelfile referencing the GGUF binary and issuing:
1ollama run aion-instruct:2.3b
If you are deploying on Windows Copilot+ hardware, Microsoft provides implementation samples via the Aion-Instruct-Preview-Sample repository using WinUI 3 and ONNX Runtime execution providers tailored for low-power NPU inference.
Google's Gemma 2 2B is one of the most capable sub-3B open weights models available. While Gemma 2 2B posts exceptional zero-shot reasoning benchmarks for its scale, it is bound by an 8,192 token context window and carries a restrictive commercial license. Aion 1.0 Instruct 2.3B offers four times the context length (32K), runs under the fully permissive MIT license, and is specifically aligned for practical desktop tooling and document workflows. Gemma remains stronger in pure academic knowledge retrieval, but Aion is substantially better suited for local file ingestion and long-form document summarization.
Meta's Llama 3.2 3B is slightly larger at 3.21B parameters and natively supports a 128K context window. In terms of raw general-purpose knowledge and complex multi-turn reasoning, Llama 3.2 3B holds an edge. However, Aion requires roughly 30% less memory bandwidth and compute, making it much easier to sustain higher tokens per second on resource-constrained laptops, edge devices, and integrated NPUs. Furthermore, Aion's native alignment with the upcoming Windows system runtime gives developers building desktop software an identical local reference environment for what will soon ship natively across hundreds of millions of consumer Windows installations.

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Microsoft model we track.

Check which of your devices can run this model and how fast.

Compare hosted per-token prices before you commit to local hardware.

Break-even math for self-hosting, renting a GPU, or paying per token.
Live status for the hosted APIs you might use instead of running this locally.