Wrivio
Get Wrivio
5 min readBy Wrivio Team

Surface Laptop Ultra, 128GB, and Running Big Models Locally

At its October 7, 2026 event, Microsoft announced the Surface Laptop Ultra, built around NVIDIA’s RTX Spark platform with up to 128GB of unified memory and the claim that it can run AI models exceeding 120 billion parameters locally. It is set to be available on October 16, 2026. For a few years, “run a frontier-sized model on a laptop” was a thing you could not do without a desktop full of GPUs. That is changing, and it is worth being clear about what changed and what did not.

What Unified Memory Actually Changes

The bottleneck for running a large model locally has never really been raw compute. It is memory. A model’s weights have to fit somewhere the processor can reach quickly, and large models need a lot of room. On a typical setup, that means fitting into a graphics card’s dedicated VRAM, which is expensive and small.

Unified memory pools system and graphics memory so the model can use a much larger shared space. 128GB is enough headroom to hold a model that would never fit in a normal laptop GPU. That is the real unlock: not that the laptop is suddenly a supercomputer, but that the memory ceiling moved up far enough to hold big weights at all. If you want the underlying relationship, why context length costs you RAM and how much VRAM for a local LLM in 2026 explain where the memory goes.

Big And Local Is Not The Same As Fast

A model fitting in memory and a model running quickly are different things. A 120B model held in unified memory will still generate tokens more slowly than a small model that fits in fast dedicated memory, because unified memory trades some speed for size. For interactive work, that tradeoff is real.

So the honest picture is that this hardware lets you run a large model locally, not that it makes running one feel instant. Whether that matters depends entirely on what you are doing. For a long research synthesis you set running and come back to, slower is fine. For rewriting an email you want polished in a second, slower is the whole problem.

Writing Does Not Need Any Of This

Here is the part the spec sheet will not tell you. Rewriting text does not need a 120B model, and it certainly does not need a 128GB laptop to run one. A 1.7B to 4B model handles editing, toning, and cleanup well, and it fits comfortably on ordinary hardware with memory to spare.

Wrivio’s Local engine runs Apache 2.0 Qwen3 models at exactly those sizes precisely because writing is not a task that rewards a giant model. It runs in-process on a normal laptop and makes zero network calls during a rewrite. The full version of this argument is in do you need a 120B model to rewrite text and small language models beat big ones for rewriting.

That is not a knock on the Surface Laptop Ultra. If your work is heavy local reasoning or running big models offline for other reasons, the memory ceiling it raises is genuinely useful. Just do not buy it for a writing workflow that a far cheaper machine already runs well.

Matching The Machine To The Job

Before:

We need the 128GB laptop so we can run AI locally for the team’s writing.

After:

Our writing use runs fine on a small local model on the machines we have. The 128GB laptop only makes sense if someone needs to run very large models offline, which is a different job. Let us decide per role, not buy the top spec by default.

The second version matches the hardware to the actual task instead of the headline. A Wrivio Context for turning a spec announcement into a buying note could say:

Rewrite these hardware notes into a short, neutral buying recommendation for my team. Separate what the machine enables from what our writing workload actually needs. Keep every figure, model size, date, and price exactly as written. Do not overstate performance or imply a capability the notes do not state.

Press Ctrl+Shift+Space, paste your notes, and check the diff so no spec got inflated. Microsoft’s own Windows blog is the place to confirm the final specs and availability; treat these details as current to October 2026. For the general sizing question, see how much RAM to run a local LLM.

Common Questions

Can the Surface Laptop Ultra really run a 120B model locally?

Microsoft says its RTX Spark design with up to 128GB of unified memory can run models exceeding 120 billion parameters locally, available from October 16, 2026; fitting a model in memory is real, though it runs slower than a small model in fast dedicated memory.

What does unified memory change for local AI?

It pools system and graphics memory into a much larger shared space, raising the ceiling on how big a model can fit, which is the main barrier to running large models on a laptop.

Do I need this hardware to use AI for writing?

No, rewriting and editing run well on a 1.7B to 4B model on ordinary hardware, so a 128GB laptop is overkill for a writing workflow and only makes sense for genuinely large local models.

Is a big local model slower than a small one?

Yes, a large model in unified memory generates tokens more slowly than a small model in fast dedicated memory, so it is better for batch work you wait on than for interactive rewriting.

Download Wrivio for Windows to run a right-sized local writing model on the laptop you already own.