The pitch fits on a napkin: the smartest AI model in the world is useless on a drone with eight gigabytes of memory. General Instinct, a two-person startup in San Francisco backed by Y Combinator's Spring 2026 batch, spent years running into that wall inside robotics projects before deciding to build the wall-remover. Its software takes the enormous models that live comfortably in a datacenter and squeezes them until they fit on the hardware a robot actually carries - fully offline, and fast enough to be useful.
The problem is more specific than "AI is big." Datacenter models are designed assuming there is always more GPU memory and bandwidth to spare. Edge devices - robots, drones, IoT sensors, an aging laptop bolted to a machine on a factory floor - assume the exact opposite. A model that works flawlessly in a demo can simply refuse to load on the chip that ships in the product. General Instinct's founders describe it plainly.
The best models never fit the hardware we actually had available.
It is the kind of sentence that only sounds obvious after someone says it. The public story of AI over the past few years has been a story of scale - more parameters, more context, more compute. Almost none of that story is about subtraction. Yet subtraction is precisely what a robotics team needs when the model that scored beautifully in the notebook has to survive on a board the size of a deck of cards, drawing a few watts, with no server to phone home to. The gap between those two worlds is where a lot of promising physical AI quietly stalls.
Section 01 / What it doesModel in, offline runtime out
The company's main product is called Instinct Edge. The interface is close to a vending machine for optimized AI: you hand it a model, a target device, and a latency budget, and it hands back an offline runtime tuned to hit that budget on the specific chip you named. Supported targets include NVIDIA Jetson boards, mobile NPUs, ARM CPUs, Apple's Neural Engine, and Qualcomm Snapdragon processors. There is no cloud call at runtime - the point is that the model runs where the device is, even with the wifi off.
Underneath sits the harder engineering. General Instinct distills and quantizes large models down to roughly three effective bits, then rebuilds the accuracy it loses along the way. The team keeps the parts of a model that always fire - routers, normalization layers, the vision pathway - at higher precision, while aggressively compressing the routed experts that only activate some of the time. Then it uses on-policy distillation to teach the shrunken model back toward the original's behavior.
The last part is the trick most compression skips. Quantizing weights down to a few bits is a well-worn idea; the trouble is that a naive squeeze also crushes the model's reasoning, and you end up with something small and dim. On-policy distillation runs the shrunken model, watches where it diverges from the original, and corrects it against its own outputs rather than a fixed dataset. The goal is a model that is not just lighter but still sharp - one that gives up disk space and VRAM without giving up the answers. General Instinct also leans on tricks like streaming rarely-used experts out of system RAM, so the device only pays for the parts of the model it is using at that moment.
Roughly a 5x shrink. The compressed artifact ends up smaller than a model a fifth its size, yet General Instinct reports it outperforming that smaller model on the MMLU-Pro and GPQA-D reasoning benchmarks. Bars scaled to the 245GB original.
Section 02 / The open coreInstinctRazor, published in the open
The compression engine is open-source, released as InstinctRazor on GitHub with the recipe and full reproduction steps. That is a deliberate choice in a field where most optimization claims arrive as screenshots. Anyone skeptical that a 122-billion-parameter model can run on a modest GPU can download the toolkit and check. The published results go past a single trophy number: on Qwen3.6-35B, the same method lifts a MATH-500 score from 81.7 to 89.2 without growing the memory footprint.
Publishing the core in the open is also a distribution strategy dressed as generosity. The people most likely to try InstinctRazor are the same engineers who would otherwise reinvent a rougher version of it on a Friday afternoon - and they are precisely the customers General Instinct wants for the paid pipeline on top. Open benchmarks that anyone can reproduce do more to earn that audience's trust than any amount of marketing copy. In a market crowded with confident claims, "here is the recipe, run it yourself" is a stronger position than it looks.
Section 03 / Why it mattersThe latency budget is the whole game
For a robot, intelligence that arrives late is not intelligence. A machine reaching for an object, a drone adjusting to wind, a camera deciding whether to stop a conveyor - all of it runs against a clock measured in milliseconds. General Instinct's demo figure speaks that language: a multimodal classifier on a Jetson Orin NX with a 111ms cold start, and every decision landing inside a 150ms budget.
The cloud does not solve this. Sending each frame to a server and waiting for an answer adds a round trip that a fast-moving machine cannot afford, and it assumes a connection that a drone over a field or a robot in a warehouse basement may not have. There is a privacy dimension too: plenty of industrial and consumer devices are not supposed to stream their camera feed to anyone's datacenter. Running the model on the device sidesteps all three problems at once - latency, connectivity, and data leaving the box - which is the whole reason on-device inference keeps coming up as the next frontier rather than a niche.
COLD START 111 ms · BUDGET 150 ms
On a Jetson Orin NX, every one of the classifier's decisions came back before the deadline. For physical AI, that reliability line matters more than a flashy average - a model that is usually fast still misses the moments that count.
Section 04 / The marketPicks and shovels for physical AI
General Instinct is not building a robot. It is building the layer that makes any robot's brain run - which, in a gold rush, is often the steadier business. As physical AI teams multiply, most of them will hit the same deployment wall the founders did, and few will want to spend months hand-writing CUDA, Metal, and ARM NEON kernels to get a model onto a chip. That months-long slog is exactly what Instinct Edge tries to collapse into a pipeline.
The company sits alongside general-purpose tooling - NVIDIA's TensorRT and Jetson stack, ONNX Runtime, Qualcomm's AI Engine, Apple Core ML, and the llama.cpp and GGUF ecosystem. Its distinction is the shape of the offer: not a library you wire together yourself, but an end-to-end path targeted to one device and one latency budget, with the compression, custom kernels, and serving pipeline bundled.
Those existing tools are powerful, but each solves a slice. A team stitching them together still owns the integration - matching a quantization scheme to a kernel to a runtime to the specific chip on the bill of materials, then keeping the whole thing fast as the model changes. General Instinct's bet is that this integration work is common enough, and painful enough, to be worth buying rather than maintaining. If physical AI grows the way its backers expect, the number of teams facing that exact chore grows with it, and the layer underneath becomes a market of its own.
Any frontier model. Any edge device.
Section 05 / The foundersTwo builders with early starts
The team is small and unusually credentialed for its size. Bill Jiao, co-founder and CEO, previously worked on Siemens' multimodal foundation model and, before any of that, secured patents on tardigrade proteins while still in high school. Guanming Wang, his co-founder, is a former Google DeepMind reinforcement-learning researcher focused on robotics who published work at ICRA at seventeen, with earlier stints tied to UCL and Tsinghua. The two have a long history of collaborating on machine-learning and vision research, and they build in person.
The biographies are worth pausing on because they explain the company's shape. This is not a pair of infrastructure engineers who wandered into robotics; it is two people who did the robotics research first, felt the deployment pain directly, and only then went looking for the layer that was missing. That order matters. The founders are building a tool for a version of themselves from a few years ago - the researcher with a good model and a board that would not run it. Startups that come from a wall the founders hit personally, more than once, tend to aim at the right problem instead of an adjacent one that is easier to demo.
Section 06 / The businessSell the layer, not the model
General Instinct's customers are the physical AI teams themselves - the people putting frontier models on robots, drones, and IoT devices and needing them to run reliably in production, offline. The business is infrastructure: deliver the optimized runtimes and the deployment pipeline, with the open-source InstinctRazor core seeding adoption among the exact engineers who would otherwise build a worse version in-house. The company is early - two people, a 2026 founding, seed-stage backing from Y Combinator - and much of the commercial detail is still ahead of it.
What is already clear is the wager. If the next wave of AI happens on devices rather than in the cloud - pushed there by latency, privacy, and plain lack of connectivity - then something has to make big models fit small chips without throwing away the intelligence. General Instinct is betting that "something" is worth building as its own company.