Microsoft: RTX Spark laptops run 284B local models that beat GPT-5 in coding

At the Surface Laptop Ultra announcement, Microsoft's Windows + Devices EVP Pavan Davuluri said RTX Spark laptops can run open-weight 284B models locally, citing DeepSeek-V4.1-Flash at 1.6-bit quantization that fits in 60GB of memory and outperforms GPT-5 in coding and reasoning benchmarks. Nvidia had earlier claimed 120B models with a one-million-token context window.

• DeepSeek-V4.1-Flash at 284B parameters fits in 60GB of memory
• Top-end Surface Laptop Ultra offers 128GB of unified RAM
• Nvidia claimed 120B models with a 1M-token context window
• Windows 11 adds a setting to allocate memory to the GPU for AI