Every agent you've built this year has the same invisible tax: a network hop and a token meter on every single tool call. For a frontier reasoning task, that's a fair trade. But most of what an agent actually does all day is small — decide which tool to call, pull three fields out of a file, route a request, summarize a page. Paying cloud prices and cloud latency for a thousand of those a day is like taking a taxi to the mailbox.
Microsoft Aion 1.0 is a bet that a lot of that work belongs on the device. It's a family of on-device small language models Microsoft announced at Build 2026 on June 2, built to run inside Windows itself — no cloud call, no per-token bill. And the piece founders should care about is the tool-calling one.
What it is#
Two models, one idea — the model lives where the work happens.
- Aion 1.0 Plan — a 14-billion-parameter reasoning and tool-calling model with a 32K-token context, meant to ship in-box on "capable" devices (a Copilot+-class NPU or a capable GPU). This is the agentic member: Microsoft built it to reason over user intent, invoke tools, manage files, and orchestrate sub-agents — the full local agent loop.
- Aion 1.0 Instruct — a smaller, efficiency-first model that runs on an ordinary CPU, so it works on far more machines than any prior in-box Windows model. Microsoft hasn't published its exact parameter count. This is the one you can actually touch today.
The through-line: an agent that reasons and calls tools on the machine, with no round-trip to a datacenter and nothing on a token meter.
Who it's for#
Founders and small teams who want a private, offline, zero-marginal-cost agent for local or desktop workloads — on-device file triage, privacy-sensitive drafting, local automation, a routing layer that decides what's worth escalating. It is not a frontier model: if you need hard multi-step reasoning or a million-token context, that's still a cloud tier's job (and if you're pricing those, we did the rent-vs-self-host math on Kimi K3 and why one tokens-per-second number lies to you).
How to start today#
You don't have to wait for the in-box rollout to try local agentic inference.
- Preview path. Install Edge Insider and reach Aion 1.0 Instruct through the Windows Copilot Runtime API — Microsoft's on-device inference surface.
- Local-server path (works now, any model). Run Microsoft Foundry Local, an OpenAI-compatible model server for Windows, macOS, and Linux that auto-selects CPU, NVIDIA CUDA, AMD DirectML, or NPU. Because the API is OpenAI-shaped, your existing SDK code just points at
localhost:
from openai import OpenAI
# Foundry Local exposes an OpenAI-compatible endpoint on the machine
client = OpenAI(base_url="http://localhost:5273/v1", api_key="not-needed")
resp = client.chat.completions.create(
model="aion-1.0-instruct", # once the weights are pulled locally
messages=[{"role": "user", "content": "Which tool handles a refund? Return JSON."}],
tools=my_tools, # tool-calling runs entirely on-device
)- Open-weights path. The Instruct open weights land on Hugging Face in July 2026 — once they're up, you can pull them into llama.cpp or MLX and run the model wherever you want, including a Mac.
What it costs#
Free. Aion ships in-box with Windows and is released as open weights, so there's no license fee and no token bill. You pay for hardware and ops — which is exactly the trade a cost-sensitive founder wants for high-volume light work: the marginal cost of the thousand-and-first tool call is zero.
The honest catches#
Three, and they matter. Plan needs a capable device — a Copilot+-class NPU or GPU — and its in-box rollout is "coming months," not shipped; Instruct is what you can run today. It's a small model: great for tool-calling and routing, not a frontier reasoner. And "on-device" means you own the ops — uptime, updates, hardware — you've swapped a token bill for a maintenance surface.
The takeaway for a team of one: stop paying frontier prices for mailbox trips. Put a local model like Aion in front of your agent as a cheap, private router — let it handle the easy, high-volume calls on the device, and escalate only the hard ones to the cloud. The token bill you save is the whole business case — and we worked out the exact volume where on-device beats a cloud API, so you can check the math against your own traffic before you commit.
