General Instinct

Summary

About

About General Instinct

Instinct Edge converts larger vision, language and multimodal models into runtimes that fit the memory, power and accelerator constraints of deployed hardware. The workflow evaluates the original model, applies quantisation, pruning or distillation and tests the compressed version directly on the target device. Supported environments include NVIDIA Jetson, Snapdragon, mobile NPUs, Arm processors and Apple Neural Engine hardware.

The customer retains the model and application while General Instinct engineers the deployment path around required latency, accuracy and offline operation. This allows a robot or drone to execute perception and decision workloads locally without a continuous cloud connection. Benchmarking on the physical device exposes thermal throttling, memory pressure and operator latency that would not appear from model-size estimates alone.

PRICE
Range 1Y
Return
High
Low
Updated
DATA: KORTHOS

Team

  • Bill Jiao - Co-founder and Chief Executive Officer
  • Guanming Wang - Co-founder and Chief Technology Officer - Develops world-action modelling and efficient robot-policy inference at General Instinct.

Research

  • General Instinct releases InstinctFlash for robot-policy inference

    2026-09-15 - Research Publication - The runtime exposes model-specific adapters through a shared Python prediction API and WebSocket serving interface. September Jetson Thor measurements reduced pi0.5 median prediction latency from 408.58 ms to 51.85 ms with FP8. The reported 33.78× LingBot-VA improvement combines FP8 with a reduced sampling schedule; it measures prediction latency rather than task-completion time.

    Source
  • General Instinct introduces InstinctRazor compression framework

    2026-06-05 - Research Publication - The pipeline preserves continuously active model components while quantising routed experts more aggressively, with on-policy distillation used to recover quality. The company’s packed Qwen3.5-122B-A10B artifact occupies about 48 GiB and fits fully on one 80 GB GPU. CPU expert offload uses roughly 48 GiB of system memory and about 7.6 GiB peak GPU memory at an 8k context; the reported offload throughput is about 46 tokens per second.

    Source

Timeline

Events

  • 2026-10-13 - Product Launch - Physical Intelligence opens Pi API beta to more robot developers

    Teams can upload their robot data, launch fine-tuning on PI-managed compute and infrastructure, and run the resulting models through hosted inference. Early users include Argon Robotics, which trained a dish-racking policy within 35 hours, and General Instinct, which trained a server-assembly policy within three days. Beta participation is by application; most showcased applications use cloud-hosted inference, with network connectivity identified as an area for improvement.

    Source
  • 2026-09-15 - Research Publication - General Instinct releases InstinctFlash for robot-policy inference

    The runtime exposes model-specific adapters through a shared Python prediction API and WebSocket serving interface. September Jetson Thor measurements reduced pi0.5 median prediction latency from 408.58 ms to 51.85 ms with FP8. The reported 33.78× LingBot-VA improvement combines FP8 with a reduced sampling schedule; it measures prediction latency rather than task-completion time.

    Source
  • 2026-06-10 - Partnership - General Instinct joins NVIDIA Inception

    The company announced its NVIDIA Inception membership on June 10, 2026. Its stated focus within the programme was edge-deployed world action models. Inception membership provides startup technical resources, training and access to hardware, software and cloud benefits.

    Source
  • 2026-06-05 - Research Publication - General Instinct introduces InstinctRazor compression framework

    The pipeline preserves continuously active model components while quantising routed experts more aggressively, with on-policy distillation used to recover quality. The company’s packed Qwen3.5-122B-A10B artifact occupies about 48 GiB and fits fully on one 80 GB GPU. CPU expert offload uses roughly 48 GiB of system memory and about 7.6 GiB peak GPU memory at an 8k context; the reported offload throughput is about 46 tokens per second.

    Source
  • 2026-05-08 - Award - General Instinct joins Y Combinator Spring 2026

    The active Spring 2026 company was founded in San Francisco in 2026. Its Instinct Edge platform distills and quantizes large models for Jetson, mobile NPU, ARM and Apple hardware. A disclosed Jetson Orin NX deployment achieved a 111-millisecond cold start and kept every decision within 150 milliseconds without cloud calls.

    Source

Relations

Relationships

  • NVIDIA

    Nvidia Inception Member - General Instinct joined NVIDIA Inception to explore edge-deployed world action models.

  • Y Combinator

    Startup Accelerator - General Instinct participated in Y Combinator Spring 2026 while building deployment infrastructure for physical AI.

  • Physical Intelligence

    Pi API Collaboration - General Instinct uses the Pi API to train server-assembly policies with reduced fine-tuning data.

Context

Shared workflows

Context tags

  • Edge AI inference - Workflow - General Instinct deploys compressed models on edge hardware.