We applied to NVIDIA's Codefest — to find out what really fits on 8 GB.

Today we applied to NVIDIA’s Open Models Codefest — fifteen teams, GPU access, and NVIDIA engineers alongside, working on the open model families. Selection lands in early August. We’re building either way, but this is a programme worth wanting.

What we make. A bolt-on brain for uncrewed ground and aerial machines. It runs the entire loop on the machine itself: speech in, an open vision-language model deciding and calling tools, speech out, motor control on the wire. On a Jetson Orin Nano 8 GB. Offline — no cloud, no GPS, no link to lean on.

That constraint is the product. Anyone can look clever with a datacentre behind them; the hard version is being useful on a board you can carry, on a battery, somewhere the network has already failed. Defence and civil protection don’t get to assume coverage, and neither do we.

Why this programme. Our whole business sits exactly where the Codefest points: open models, small enough to run at the edge, doing real work. NVIDIA publishes models in the weight class our board can actually hold, and which one to run is not an academic question for us — it’s the difference between a machine that reads a scene correctly and one that drives into a wall.

We can’t answer it properly on our own hardware. That’s what we want from this.

What we’d do with it. Benchmark NVIDIA’s open models against the one we ship today — on our metric, not a leaderboard’s. We grade navigation on collision rate, not trajectory error, because a path can score beautifully on trajectory error and still go through a wall. A model that reasons elegantly about a photograph and clips a doorframe is no use to anybody.

We’d also want to know what the runtime costs us. A model that drops into the stack we already run is a different proposition from one that needs the stack rebuilt around it — and on an 8 GB board that decision is measured in megabytes, not preferences.

Where we stay honest. In R&D we fine-tune an open model into a navigation planner, and that adapter is not deployed. What ships on the robot today is the base model. When the fine-tune earns its place on hardware, we’ll say so — with the numbers, including the ones that don’t go our way. That’s been the habit on our public model card, and we’re not about to change it for a competitive application.

Fifteen places, and we don’t know yet. What we do know is the question we want answered — and that it’s the same one every customer eventually asks: what can this machine actually do when it’s on its own?

If you’re building autonomy for the places the network can’t reach, talk to us.