Launch HN: General Instinct (YC P26) – Frontier models on edge devices
245 GB compressed to 48 GB. General Instinct just made frontier models actually fit on edge devices.
Why it matters
A YC startup has cracked a critical constraint for AI deployment: running state-of-the-art models on resource-limited hardware (robots, edge devices) without losing capability. This matters because most frontier models are datacenter-optimized; getting them to work locally could unlock a new class of autonomous systems.
The key facts
7 to knowQwen3.5-122B-A10B (245 GB BF16 MoE) compressed to 48 GiB GGUF
Compressed model smaller than Gemma-4-26B while outperforming on MMLU-Pro and GPQA-D benchmarks
Peak VRAM usage 7.6–8 GB with 8k context window in small GPU configuration
Experts can be streamed from system RAM for reduced GPU memory
Open-sourced InstinctRazor tool for model compression
Approach: preserves always-active components (router, norms, Gated-DeltaNet/SSM layers, vision pathway), quantizes routed experts aggressively, uses on-policy distillation for capability recovery
Founded by Guanming and Bill, YC P26 cohort, background in robotics
Go to the source
Hacker Newsnews.ycombinator.com
Publisher excerpt: Hey HN, Guanming and Bill here from General Instinct ( After years of working in robotics, we kept running into the same problem: the best models never fit the hardware we actually had available. The models that performed best were usually designed around datacenter assumptions: large GPUs, lots of…