档案库 · AI 与模型 · 技术决策 · 2026
通用本能押注前沿模型必须适配边缘设备——245GB压缩至8GB
两人的YC P26初创公司将数据中心级前沿模型压缩为机器人和无人机的离线运行时,并开源其压缩引擎
General Instinct
做的是什么生意
General Instinct's Instinct Edge takes a frontier model, a target edge device, and a latency budget, and returns an offline runtime tuned to that hardware — no cloud call at runtime.
起因
Bill Jiao and Guanming Wang, a two-person team in San Francisco, founded General Instinct in 2026 after years of robotics and ML research — Jiao previously worked on Siemens' multimodal foundation model, Wang was a Google DeepMind reinforcement-learning researcher. Their repeated experience: the best models never fit the hardware actually available on robots, drones, and factory devices.
经过
Instinct Edge distills and quantizes large models to roughly three effective bits, keeping always-active parts such as routers, normalization layers, and vision pathways at higher precision while compressing routed experts aggressively, then uses on-policy distillation to recover lost reasoning. Published results include compressing Qwen3.5-122B-A10B from a roughly 245GB BF16 model to a 48GiB GGUF that runs in 7.6–8GB peak VRAM, and a multimodal classifier on Jetson Orin NX with a 111ms cold start delivering every decision inside a 150ms budget. The compression engine is open source as InstinctRazor.
结果
As of the YesPress profile, General Instinct is still early: two people, YC Spring 2026 backing, and no disclosed customers or revenue. Its Launch HN drew sharp questions from the quantization community about benchmarks against Unsloth and other 3-bit methods, which the founders answered by pointing to reproducible open pipelines.
背景
General Instinct由Bill Jiao和Guanming Wang于2026年创立,两位机器人学和机器学习研究员反复遇到同一困扰:最好的AI模型是为数据中心假设设计的,而机器人、无人机和边缘设备具有相反的约束。他们的赌注是物理AI需要设备本地运行前沿级模型,完全离线。
该公司的Instinct Edge产品将该赌注转化为流水线:给定一个模型、一个目标芯片和一个延迟预算,返回离线运行时。在底层,它将模型量化到约三个有效比特,同时保护始终活跃的组件,然后使用同策略蒸馏恢复能力。头条结果是245GB的Qwen3.5-122B MoE模型压缩至48GB并可在约8GB显存内运行,且每个分类器决策在Jetson Orin NX上落在150ms预算内。
这家两人初创公司还开源了其压缩引擎InstinctRazor,作为一种刻意的信任与分发策略——怀疑者可以自行重现其声称。General Instinct押注随着物理AI的增长,使前沿模型适配小芯片的集成工作将自成市场,销售该层比构建另一个机器人更好。
这件事要成立,得有什么
- 两位创始人在机器人领域亲身遭遇部署瓶颈,因此公司围绕他们实际遇到的问题而非容易演示的相邻问题构建
- 同策略蒸馏加选择性精度比标准3-bit方法压缩得更狠,同时恢复推理能力,解决通常的小而笨权衡
- 开源InstinctRazor为持怀疑态度的工程师提供了可复现的证据,并在付费流水线瞄准的客户中播下采用种子
- 延迟、连接性和隐私推动推理转向设备端,因此物理AI的部署层市场潜力增长
可借鉴之处
创始人反复遇到的瓶颈是一个可靠的指南针:销售缺失的层,让任何机器人都能运行前沿大脑——并开源核心以在炒作泛滥的市场中赢得信任。
后续进展
截至2026年7月30日,General Instinct仍然是一家由YC支持的旧金山两人初创公司,未披露客户或营收。其公开产物包括开源的InstinctRazor工具包、关于亚4位MoE压缩的技术博客,以及面向Jetson、移动NPU、ARM CPU、Apple Neural Engine和Snapdragon的Instinct Edge产品。下一步是将发表的结果转化为付费的物理AI部署,同时抵御TensorRT和llama.cpp等工具及激进的社区量化。
资料来源
- Launch HN: General Instinct (YC P26) – Frontier models on edge devices
- General Instinct: Frontier AI Models on Any Edge Device
发现哪里写错了?告诉我们。
轮到你了
你刚读完一家。说说你在做什么,看看谁在赌同一件事。
免费账号 · 3 次免费提问 · 不用绑卡