In partnership with
Microsoft's October 7 Windows announcement describes a shift in where personal-computer AI work can run. The company is building a hybrid system that sends some tasks to local models and others to cloud services. It also announced Windows computers built around NVIDIA's RTX Spark platform, including models from several manufacturers. The technical question is not simply how large a model can fit. It is how software decides where a task runs, what data it can use, and how reliably the device completes it.
Take the prompts our creative team actually uses
The key to maintaining brand quality at scale? Mastering how to communicate with AI models and embedding them into your creative strategy.
Join our honest discussion with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready to run prompts.
Microsoft says one Surface Laptop Ultra configuration has up to 128 gigabytes of unified memory. The company says it can run models exceeding 120 billion parameters locally. That is a vendor statement about a particular configuration, not an independent demonstration of useful performance. Parameter count describes a model's scale, but it does not establish accuracy, latency, energy use, or suitability for a given task. Microsoft's post also describes a 137-billion-parameter coding model running at 3-bit precision. The company says that approach reduces model size nearly 80 percent while preserving coding quality. The blog does not provide an independent quality test for that statement.
The wider announcement matters because the software design treats local and remote inference as complementary. Local execution may reduce the need to send some prompts or files to a remote service. Cloud access can still be necessary when a device lacks the required model, memory, or compute capacity. Microsoft's planned Copilot features would use local context and models, then combine them with cloud intelligence when needed. The company says these features will arrive gradually, with timing varying by device, market, and silicon platform.
200+ Proven Ways to Make Money With AI in 2026
The next wave of millionaires will be people who figured out how to make AI work for them.
The window to get ahead is still open. But not for long.
Here are 200+ proven ways to make money with AI in 2026.
Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.
There is a separate systems issue: agents can act on files and tools, so model location alone does not define privacy or safety. Microsoft says its Execution Containers are generally available on Windows 11 and can enforce file and network permissions at runtime. That is a control mechanism, not proof that every agent is harmless or every configuration is secure. Effective protection also depends on the policies an administrator sets, the software's implementation, and the actions allowed inside the boundary.
Microsoft and NVIDIA also publish performance comparisons against a preproduction Apple laptop. The stated tests use particular models, memory configurations, software, image settings, and prompt lengths. Microsoft says its own testing measured one coding task, while NVIDIA tested image and video generation. These are company-controlled comparisons, not independent evaluations across ordinary applications. Results may change with device configuration and software, so the headline multipliers should not be read as general speed rankings.
The practical unknowns remain substantial. Public information does not yet establish local battery life under demanding workloads. It also does not show heat levels or consistency across applications. A task's ability to remain entirely offline is another open question when users expect local processing. The blog's own availability notes say features roll out over time and may vary by device and region. Shipping dates for announced machines are not evidence that planned software capabilities are already present.
For developers and organizations, the shift is architectural rather than a simple replacement of cloud computing. They will need to test model quality, update flows, memory pressure, access controls, and the points where data leaves a device. Procurement comparisons should use the same tasks and settings on the actual configurations under consideration. Microsoft's announcement makes the local-cloud boundary more visible, but independent, workload-specific testing will determine what the design delivers. That is the useful question to revisit as systems reach users.
Some teams never seem to stop moving. They're on Attio, the agentic CRM.
Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.
With Attio, you’ll get:
Leads automatically prioritised and routed to the right rep
Expansion and risk signals caught the moment they land
Follow-ups written in your voice, already there when you arrive
Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?
