Aman Sachan compresses DeepSeek-Coder-V2-Lite to 4.5 GB for laptop GPUs

Developer Aman Sachan has released a compressed build of the open-weights DeepSeek-Coder-V2-Lite-Instruct model packaged with MAPLE ^[1]. The project focuses on trimming the footprint of the instruction-tuned coding assistant so that it fits completely inside the memory of modest consumer computers ^[1].

According to the project release notes, the packaged MAPLE container size is 4.50 GB ^[1]. Compressing the model to this file size allows developers to load and retain the entire weight structure directly within local video hardware rather than relying on host system memory ^[1].

Benchmark measurements reported by Sachan show that the compressed setup maintains substantial throughput on everyday portable hardware ^[1]. Tested on an RTX 4050 laptop GPU configured with 6 GB of memory, the model achieves inference processing speeds between 65.6 and 76.3 tokens per second ^[1].

Deploying advanced coding models on mobile workstations usually forces builders to make severe performance trade-offs, particularly when system memory offloading slows generation speeds. By keeping the container footprint at 4.50 GB, this MAPLE build demonstrates that capable coding models can generate rapid completions on mainstream mobile chips with only 6 GB of dedicated video memory ^[1].