![\"Microsoft's](\"https:\/\/www.newsbeep.com\/il\/wp-content\/uploads\/2026\/10\/decode.png\")
<\/p>\n<\/p>\n
Microsoft\u2019s local MAI-Code-1.1-Flash generates roughly 60 tokens per second with short prompts on Surface Laptop Ultra, versus just under 40 at a 256K-token prompt. The figures come from Microsoft\u2019s testing on a synthetic code-generation workload. Image: Microsoft<\/p>\n
2026 saw the rise, and quick fall, of tokenmaxxing, the practice of enterprise companies treating AI spend volume as a yardstick for productivity. This summer, Microsoft put its divisions on AI token budgets and told engineers that \u201ctokenmaxxing is not what we are optimizing for.<\/a>\u201d<\/p>\n Now, in its Oct. 7 announcements<\/a>, it offered customers a way to avoid the same sticker shock by moving more of the agents\u2019 work onto the PC. It imagines always-on mini PCs running agents, and shared Nvidia DGX Stations, desk-side AI workstations, serving 32 or more people at once. The cloud stays available for work that needs it.<\/p>\n Coding is an early test of Microsoft\u2019s proposition. The company is bringing its MAI-Code-1.1-Flash<\/a> model to local coding workflows, where an agent can write code, run tests and revise its work through repeated model calls. Running those calls locally removes a per-inference charge.<\/p>\n The catch with local inference has been capability: models small enough to fit in a few dozen gigabytes have generally trailed hosted models on serious coding work. Microsoft\u2019s numbers suggest that gap is narrowing.<\/p>\n In Microsoft\u2019s testing, the on-device version of MAI-Code-1.1-Flash resolved 70.8% of tasks on SWE-bench Verified, a benchmark built from GitHub issues, compared with 72.6% for its higher-precision cloud variant. The quantized model occupies about 53GB, with peak memory use reaching 75.5GB at its full 256K-token context. Microsoft compared it against another open model, a quantized version of OpenAI\u2019s gpt-oss-120b, rather than frontier models.<\/p>\n Microsoft\u2019s new Surface Laptop Ultra<\/a> uses Nvidia RTX Spark hardware and offers up to 128GB of shared memory. Preorders opened Wednesday at a starting price of $2,599, with availability scheduled for Oct. 16.<\/p>\n GitHub\u2019s HydraFusion is designed to route work between local and cloud models within a coding session, with an experimental preview due later this month in the GitHub Copilot app, GitHub Copilot CLI and Visual Studio Code. The intent is for the laptop to handle the repetitive loop of edits and test runs, while larger models remain on call for harder problems. Developers can also select the local model explicitly.<\/p>\n A coding agent left to keep trying also needs limits on what it can touch. Microsoft made its Execution Containers<\/a> generally available Wednesday, providing policy-driven containment for local agent processes. Entra attribution and Agent 365 and Intune controls for local agents remain forthcoming; microVM isolation is experimental. And local inference does not make an entire session offline: remote tools and separate cloud requests can still send work beyond the PC.<\/p>\n