# General Instinct

> General Instinct is a Y Combinator (Spring 2026) startup building inference infrastructure for physical AI. Its software distills, quantizes, and deploys large frontier models onto resource-constrained edge hardware - drones, robots, IoT devices, and old PCs - fully offline. The flagship product, Instinct Edge, takes a model, a target device, and a latency budget and returns an optimized offline runtime that hits that budget on hardware like Jetson, mobile NPUs, ARM CPUs, and Apple Neural Engine. The company also ships an open-source compression toolkit, InstinctRazor.

- **Founded:** 2026
- **Headquarters:** San Francisco, California, United States
- **Founders:** Bill Jiao (Co-Founder & CEO), Guanming Wang (Co-Founder)
- **Team size:** 2
- **Products:** Instinct Edge, InstinctRazor
- **Notable:** Compressed Qwen3.5-122B-A10B (245GB) to a ~48GB GGUF - smaller than Gemma-4-26B while outperforming it on MMLU-Pro and GPQA-D., Ran a compressed 122B-class model with ~7.6-8GB peak VRAM at an 8k context window., Demonstrated a multimodal classifier on Jetson Orin NX with 111ms cold start and 100% of decisions inside a 150ms budget.

## Products & services

- **Instinct Edge** — A no-code pipeline that takes a model, a target device, and a latency budget and returns an offline runtime optimized to hit that budget on hardware like Jetson, mobile NPUs, ARM CPUs, Apple Neural Engine, and Snapdragon.
- **InstinctRazor** — Open-source model compression toolkit that quantizes any model sub-4-bit and recovers accuracy loss with on-policy distillation. Published a Qwen3.5-122B-A10B compressed to a ~48GB GGUF that runs on a single small GPU.

## Achievements

- Compressed Qwen3.5-122B-A10B (245GB) to a ~48GB GGUF - smaller than Gemma-4-26B while outperforming it on MMLU-Pro and GPQA-D.
- Ran a compressed 122B-class model with ~7.6-8GB peak VRAM at an 8k context window.
- Demonstrated a multimodal classifier on Jetson Orin NX with 111ms cold start and 100% of decisions inside a 150ms budget.
- Improved Qwen3.6-35B MATH-500 from 81.7 to 89.2 at the same memory footprint via InstinctRazor.
- Accepted into Y Combinator (Spring 2026 batch).

## Latest updates

- **2026-03** — Joined Y Combinator's Spring 2026 (P26) batch and closed initial seed funding.
- **2026-03** — Launched on Hacker News: frontier models on edge devices, with open-source InstinctRazor.

## Links

- Website: https://general-instinct.com
- LinkedIn: http://www.linkedin.com/company/general-instinct
- Twitter/X: https://twitter.com/BillJiao930
- GitHub: https://github.com/General-Instinct/InstinctRazor

---

Profile page: https://yespress.io/general-instinct-yc-p26
Published by YesPress — https://yespress.io
Last updated: 2026-07-31