---
format: "aidr-story-markdown/v1"
id: "c4c31b79a93293bf18ea9227fd360704171d2e0a384a91d8de1e24a0f77590a7"
canonical_url: "https://aidr.today/c4c31b79?lang=en"
title: "Microsoft Routes Copilot Tasks Locally to End Inference Charges"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-07T19:39:09.000Z"
category: "Tools"
topics: ["microsoft","inference","devtools"]
source_urls: ["https://marketbrief.now/ai/microsoft-routes-copilot-tasks-locally-to-end-inference-charges-e2ef6704","https://x.com/satyanadella/status/2107898018112647313","https://x.com/msdev/status/2107896901970665488","https://x.com/mweinbach/status/2107890239360676148","https://x.com/abidlabs/status/2107915489670508830","https://x.com/pavandavuluri/status/2107910939555271097","https://x.com/rohanpaul_ai/status/2107914219408736463","https://x.com/PhotonCap/status/2107909946826453124"]
summary: "Microsoft introduced HydraFusion, a routing system that distributes AI tasks between on-device and cloud models within the Windows ecosystem to reduce operational costs. GitHub Copilot now routes routine coding work from Anthropic's Claude Haiku 4.5 to the locally running MAI-Code-1.1 Flash, a 137B parameter model with a 256K token context window, with on-device requests carrying no inference charges for users. The software depends on Nvidia RTX Spark hardware and Microsoft Execution Containers (MXC), now generally available on Windows 11, isolates agent activity to designated files and network locations for security."
---

# Microsoft Routes Copilot Tasks Locally to End Inference Charges

> [Open the canonical story](<https://aidr.today/c4c31b79?lang=en>)

**Published:** 2026-10-07T19:39:09.000Z
**Category:** Tools
**Topics:** microsoft, inference, devtools

## Summary

Microsoft introduced HydraFusion, a routing system that distributes AI tasks between on\-device and cloud models within the Windows ecosystem to reduce operational costs\. GitHub Copilot now routes routine coding work from Anthropic's Claude Haiku 4\.5 to the locally running MAI\-Code\-1\.1 Flash, a 137B parameter model with a 256K token context window, with on\-device requests carrying no inference charges for users\. The software depends on Nvidia RTX Spark hardware and Microsoft Execution Containers \(MXC\), now generally available on Windows 11, isolates agent activity to designated files and network locations for security\.

## Sources

- [Story source](<https://marketbrief.now/ai/microsoft-routes-copilot-tasks-locally-to-end-inference-charges-e2ef6704>)
- [Story source](<https://x.com/satyanadella/status/2107898018112647313>)
- [Story source](<https://x.com/msdev/status/2107896901970665488>)
- [Supporting source](<https://x.com/mweinbach/status/2107890239360676148>)
- [Supporting source](<https://x.com/abidlabs/status/2107915489670508830>)
- [Supporting source](<https://x.com/pavandavuluri/status/2107910939555271097>)
- [Supporting source](<https://x.com/rohanpaul_ai/status/2107914219408736463>)
- [Supporting source](<https://x.com/PhotonCap/status/2107909946826453124>)