Surface Laptop Ultra: 128GB memory sets the limit for local AI models
The Surface Laptop Ultra starts at $2,599 and offers up to 128GB shared between the processor and graphics. DeepSeek V4 Flash needs about 142GB for its 4 bit weights, while privacy and offline access make the case for running smaller models locally.
DeepSeek V4 Flash, the largest open model in NVIDIA's list for RTX Spark, has 284 billion parameters, and at 4 bit precision its weights alone come to about 142GB. The Surface Laptop Ultra, which Microsoft opened for pre order on Wednesday from $2,599 with availability from 16 October, offers up to 128GB of memory shared between CPU and GPU. The model does not fit there at 4 bit.
Memory, rather than compute, sets which models fit. The laptop's usable capacity depends on configuration and workload, as Microsoft's footnote makes clear. Cloud tokens for small models cost little enough that savings alone do not make the case for local use. The stronger reasons are keeping data on the device and working without a connection. Those are real benefits, but neither is a cost saving.
Microsoft's footnote says memory varies by configuration and that the GPU can address less than the total, depending on configuration and workload. The $2,599 price is a starting price; the post does not say which configuration has 128GB, so that figure is a maximum. The companion Surface RTX Spark Dev Box costs $5,999, ships in November and is listed with 128GB.
We calculate weight memory by multiplying a model's parameter count by the bits per weight and dividing by eight. At 4 bit, a model with 30 billion parameters needs 15GB. That covers the 30 billion parameter Nemotron 3.5 Lightning and Meta Muse Glimmer that NVIDIA lists for RTX Spark. Microsoft says the laptop can run models above 120 billion parameters locally; at 4 bit, their weights need 60GB, leaving room for the working memory of a long conversation. DeepSeek V4 Flash needs 142GB for its 4 bit weights, before file overhead.
Fitting 284 billion parameters into 128GB takes fewer than 3.6 bits per weight, and the GPU addressable share is smaller than the total, so the real threshold sits lower still. NVIDIA's own wording is that the model can run locally on a cluster of two DGX Spark machines, each with up to 128GB, which signals that the vendor does not expect it to fit on one box at full precision. A laptop owner who wants that model has to compress it hard and accept whatever that does to the answers.
A model can fit and still leave too little memory for other work. At 4 bit, a model with 120 billion parameters leaves about 68GB of a 128GB machine for the operating system, a long conversation's context and any other model running beside it. The model with 284 billion parameters leaves no room. Long context agents use more memory, and Microsoft's footnote that the GPU can address less than the total narrows the margin further.
DeepSeek V4 Flash activates only 13 billion of its 284 billion parameters for each token, which reduces the arithmetic needed and can improve speed. It does not reduce the memory required: any expert in the network might be needed for the next token, so all the weights must remain in memory. Sparse activation lowers compute demand but leaves storage demand unchanged.
Microsoft's petaflop figure is a separate matter. The post gives up to 1 petaflop of AI performance, and its footnote ties that number to theoretical FP4 performance using the sparsity feature. It describes peak arithmetic speed and says nothing about how big a model the machine can hold, so it should be read apart from the 120 billion parameter claim.
Cost makes a weak case for the hardware. Anthropic's Haiku 5.5 and OpenAI's Luna each charge $0.50 per million output tokens. At that rate, the $5,999 Dev Box, the only machine Microsoft lists with 128GB, equals about 12 billion output tokens. An open model with 120 billion parameters is not comparable to a small cloud model, so the Sonnet class at $10 per million output tokens is a fairer benchmark. The same $5,999 would buy about 600 million tokens there. A developer would need to generate that much locally for the machine to pay for itself on token costs alone.
The strongest case is for buyers for whom tokens are not the point. The relevant question is whether the cloud is allowed or reachable at all, whatever it costs. Microsoft's post on hybrid intelligence describes agents running locally or in the cloud according to the task. That fits this product better than a savings story. Routing is the product, and the laptop is its local half.
Weight memory, rather than parameter count alone, sets the practical ceiling. A model with 30 billion parameters at 15GB leaves most of the machine free, and a model with 120 billion parameters at 60GB still leaves room. The Dev Box has the same 128GB limit, so its higher price buys no extra headroom for DeepSeek V4 Flash. A model requiring 142GB at usable precision is a platform demonstration; daily users need enough memory to hold it or must compress it and accept the effect on its answers. Buyers whose data cannot leave the building or who work without a network remain the clearest audience, though the product is narrower than its headline suggests.