The Numbers Behind Osiris Compute

Anyone can say their system is fast and private. We would rather show you numbers you can check, including the ones that are not flattering. Everything below is measured on our own hardware or proven with code you can read.

One small device has a hard ceiling

On July 17, 2026 we ran a fresh benchmark on a Jetson Orin Nano, an eight gigabyte edge board of the kind an ordinary person might own. A small language model runs comfortably on it:

Then we asked the same board to load a four billion parameter model. It failed, immediately and honestly, with the message every edge developer knows: cudaMalloc failed: out of memory. The board cannot allocate the buffer. A 4B model does not fit on one small device. That ceiling is not a detail. It is the entire reason the next part exists.

Split the model, and check that it still gives the same answer

The obvious worry with spreading a model across several devices is that you end up with a slightly different, quietly worse model. We tested that worry directly rather than waving it away. We took a real pretrained decoder, distilgpt2, the exact artifact the in-browser runtime loads, cut it into stages so each device owns some of the layers, and generated text through the chain.

The split model produced token for token identical output to the unsplit model. The worst absolute difference in the raw logits over twelve decode steps was 0.000. Not close. Identical. The cut is mathematically clean, because the only thing crossing the boundary between devices is the hidden state, the key and value caches stay local to each stage, and nothing is recomputed. The code that proves it is available.

The honest part, where it is slow and why

Splitting a model buys capacity, not speed for a single stream. The stages run in sequence, so spreading one conversation across a circle does not add up to more tokens per second for that conversation. A large model spread across a circle runs on the order of one token per second for a single stream. And it is latency bound. Every token pays for the network hops between devices, so throughput falls as round trip time rises. On a wired local network, where a round trip is about three milliseconds, that cost is small. Over the open internet, where it is tens to hundreds of milliseconds, it is not. This is physics, not a bug, and it is exactly why the configuration we push is a private grid on a local network, or a dedicated box on site.

This matches the published research

We are not reporting numbers that only we see. Petals, the BigScience collaborative inference system, serves a 176 billion parameter model across volunteer hardware in a BitTorrent style network at roughly one step per second, the same regime our circle lands in. And a 2026 ACM Computing Surveys review of edge LLM inference notes that these systems are communication coupled and that typical edge deployments struggle to exceed two tokens per second. Our measurements sit right where the literature says they should.

We publish the proof and the ceiling because that is what makes the rest of it believable. The claim is not that this is faster than a data center. The claim is that it is exact, private, and yours, and now you have seen the receipts.