Kimi K3 pruned to 478GB by cutting multi-language, keeping English smarts

A viral community effort has shown that large MoE (mixture-of-experts) models can be dramatically slimmed down by removing multilingual components without sacrificing English-language intelligence. The developer hellohazine took Kimi K3, a 711GB open-source model, and carved out everything except the English-language pathways, producing a 478GB IQ2-XXS quantized GGUF called Kimi-K3-REAP-512GB-GGUF. The result is a 233GB size reduction—roughly 33% smaller—while retaining the model's high-level reasoning and coding capabilities.

Early testing added nuance to the story. Using SWE-Lancer benchmark tasks, the standard Kimi K3 2-bit model failed to solve all three tasks when streamed from SSD due to memory constraints, while the pruned 478GB variant (reap576_iq2xxs) successfully solved the same tasks. The tester noted their environment may have caused the failures, but highlighted a "micro-possibility" that pruning the expert layer could actually improve coding performance. The community sees this as a blueprint for future model optimization—the same technique could be applied to Qwen MAX 3.8 or DeepSeek V4 Flash, making frontier-scale AI more accessible to hobbyists with limited VRAM.