RTX Spark Laptops Pave The Way To Running 284B AI Models Locally That Outperform GPT-5 In Coding And Reasoning Benchmarks, According To Microsoft Executive

NVIDIA claimed during its RTX Spark presentation that laptops equipped with this SoC could run 120-billion-parameter models with a one-million Context Window. During the Surface Laptop Ultra announcement, a Microsoft executive had an even better surprise for those who’ll upgrade to the top-end notebooks down the road; these machines are able to run open-weight 284B models that outperform OpenAI’s GPT-5 in coding and reasoning benchmarks.

DeepSeek-V4.1-Flash running at a 1.6-bit quantization can run completely locally on RTX Spark laptops, according to a Microsoft executive

During the software giant’s keynote, Pavan Davuluri, Executive Vice President, Windows + Devices, was actively speaking about how open-weight models aren’t just rivaling frontier models, but with RTX Spark’s unified memory architecture, it’s possible to run 284B LLMs like DeepSeek-V4.1-Flash completely on-device.

In the presentation, the Microsoft executive says that this model can fit into 60GB of memory, so if you’ve purchased the top-end version of the Surface Laptop Ultra with 128GB of unified RAM, you’ll have plenty of available memory to substantially increase that Context Window. The launch of RTX Spark has also materialized a new feature in Windows 11, where users can determine how much memory to allocate to the GPU to tackle AI workloads.

What’s interesting about the claim made by Pavan Davuluri is that the 284B 1.6-bit version of DeepSeek-V4.1-Flash can rival GPT-5 in coding and reasoning benchmarks. Keep in mind that before GPT-6 Astra arrived, GPT-5 was the most advanced model, and to witness that an LLM that can fit entirely in video memory can zip past it in certain benchmarks is an incredible future concerning on-device AI.

Of course, it’s important to remember that claims and real-world benchmark results don’t share the same story, so we’ll have to wait for the first commercial batches to ship out and see how far these 284B models can take us.

News Source: Rohan Paul

Follow Wccftech on Google to get more of our news coverage in your feeds.