General Instinct, co-founded by Guanming and Bill, has announced the launch of their new frontier model aimed at addressing the challenges of deploying AI models on edge devices. After years in robotics, they recognized that top-performing models were often designed with datacenter constraints in mind, making them impractical for physical systems. The company has open-sourced InstinctRazor, a tool that compresses the Qwen3.5-122B-A10B model from 245 GB to a 48 GiB configuration, outperforming competitors like Gemma-4-26B-A4B on benchmarks such as MMLU-Pro and GPQA-D. The model's design allows it to function with a smaller GPU configuration, utilizing on-policy distillation to maintain performance despite aggressive quantization. General Instinct seeks feedback from users deploying models on robots and edge devices, inquiring about the challenges faced in production settings.
General Instinct Launches Edge-Compatible Frontier Models for Robotics
More Articles From This Day
Google Announces New AI Tools and Models Including Nano Banana and Co-Scientist
This week, Google introduced several AI tools and models, including the general availability of Nano Banana 2 and Nano Banana Pro via the Gemini Enterprise Agent Platform and Gemini API. The new Co-Scientist system is designed for structured scientific thinking, generating and refining hypotheses for complex problems. Additionally, dreambeans curates personalized daily topic collections based on connected Google apps. The unified encoder-free model, Gemma 4 12B, offers powerful multimodal intelligence for offline use, while Gemma 4 models have been optimized with Quantization-Aware Training to improve memory efficiency and performance. Google Magenta also launched RealTime 2, an open-weights live music model playable with MIDI keyboards and text prompts.