General Instinct open-sources InstinctRazor
General Instinct publicly launched InstinctRazor as an Apache-2.0 framework for sub-four-bit model quantization and on-policy recovery.
Evidence notes
- The framework quantizes Hugging Face models below four bits and recovers lost quality through on-policy distillation.
- Its published example compressed Qwen3.5-122B-A10B into an approximately 47-gigabyte GGUF artifact.
- The deployment path supports a single 80-gigabyte GPU or expert offload with about eight gigabytes of active GPU memory.
Company context
General Instinct builds an edge deployment layer for physical-AI systems. The company focuses on distillation, quantization, and deployment workflows that make large frontier models fit on constrained robot, drone, and edge-device hardware.