AmSach/DeepSeek-Coder-V2-Lite-Instruct-MAPLE-Sub5GB-GGUF: 4.5GB MoE Model for 6GB VRAM GPUs

Aman Sachan has released the Manifold-Activated Pruning of Latent Experts (MAPLE) Edition of DeepSeek-Coder-V2-Lite-Instruct, compressing the 15.7B parameter mixture-of-experts model down to 4.50 GB. The container fits entirely into VRAM on a 6GB consumer laptop GPU without offloading to the CPU, delivering a generation speed of 65.6 to 76.3 tokens per second. Open article →