Guanming

General Instinct, a robotics-focused company, addresses the challenge of deploying high-performance AI models on edge hardware by compressing a 245 GB BF16 MoE model (Qwen3.5-122B-A10B) into a 48 GiB GGUF format. Their method preserves critical components while aggressively quantizing routed experts and uses on-policy distillation to recover lost capability. The open-source tool InstinctRazor and a blog post detail their approach.