June 5, 2026
Big AI, tiny device, giant comment war
Launch HN: General Instinct (YC P26) – Frontier models on edge devices
They shrank a giant AI to fit tiny machines, and the comments instantly turned into a nerd fight
TLDR: General Instinct says it can shrink a massive AI so it runs on small devices like robots instead of giant servers. Hacker News loved the ambition but instantly argued over whether the comparison was fair, whether the tests proved anything, and whether this is even the right kind of AI for small hardware.
Startup General Instinct showed up on Hacker News with a big claim: it squeezed a huge AI model that normally needs server-class hardware into something small enough to run on everyday devices, even with only about 8 GB of graphics memory in one setup. In plain English, they’re trying to make powerful AI work on robots and other real-world machines that can’t rely on giant data centers. Cool demo? Yes. Quiet applause? Absolutely not.
The real show was in the comments, where the community immediately split into camps. One side basically said, “Impressive, but are we solving the right problem?” A skeptical commenter was baffled that anyone would target this style of AI for edge devices at all, arguing it’s almost built to be awkward on small hardware. Another accused the launch of comparing itself to the wrong rival, calling the benchmark framing “misleading” and suggesting stronger alternatives already exist. Ouch.
Then came the classic Hacker News energy: less hype, more interrogation. People wanted proof about what actually caused the performance gains, especially whether the team’s extra training trick really mattered or whether similar community-loved methods already do the job. One commenter also dragged the benchmark choices, basically saying some tests are so overused they barely reveal anything anymore. And for extra internet flavor, someone tossed in a random YouTube PR suggestion, because no launch thread is complete without unsolicited media strategy. The vibe: fascinated, suspicious, and very ready to argue.
Key Points
- •General Instinct says it was founded after repeated experience in robotics with frontier models that did not fit typical edge hardware constraints.
- •The startup recently open-sourced InstinctRazor as part of its effort to make large frontier models practical on edge devices.
- •The company says it compressed Qwen3.5-122B-A10B, a roughly 245 GB BF16 MoE model, into a 48 GiB GGUF model.
- •General Instinct states that the compressed model is smaller than Gemma-4-26B-A4B while outperforming it on benchmarks including MMLU-Pro and GPQA-D.
- •The post says the model can run in a small-GPU mode by streaming experts from system RAM, using about 7.6 to 8 GB of VRAM at an 8k context window.