Two PCs, two GPUs, zero cloud
Two machines on one LAN, two DeepSeek Harness agents, both fed by 1bit.MONSTER. 35B at 166 tok/s on a Radeon 8060S, 8B at 96 tok/s on an RX 9070 XT, plus Qwen3-VL vision. Six engine bugs fixed along the way, and the methodology that caught them.
The two-node fleet
Two home PCs, one flat network. Each machine runs a DeepSeek Harness agent, and each agent is fed by a 1bit.MONSTER binary on the machine in front of it. No orchestration server, no cloud endpoint, nothing between the agent and the engine except a LAN socket.
That is the whole architecture. The interesting part is what the hardware does with it.
The hardware
| Node | GPU | Model | Throughput |
|---|---|---|---|
| A | Radeon 8060S | 35B | 166 tok/s |
| B | RX 9070 XT | 8B | 96 tok/s |
Node A carries the heavy job: a 35B model at 166 tok/s on the Radeon 8060S. Node B runs a lighter 8B at 96 tok/s on the RX 9070 XT. Vision turns go through Qwen3-VL on the same hardware. One engine drives all of it; the only thing that changes between the two machines is the model that gets loaded.
Six bugs, one method
A two-node fleet is a good bug-finder, because it doubles the surface and the failure is never where you expect it. Six engine bugs surfaced across the run, and each one was fixed the same way: reproduce with a failing case before touching a line, find the smallest change that flips the case, keep the test, and record the before/after number.
The discipline is the one the rest of the site runs on. If a number cannot be reproduced on hardware, it is quarantined, not shipped. Every fix here carries its own test, so the bug stays dead.
Why zero cloud matters
The point of a fleet at home is not that it is cheap, though it is. It is that the whole stack is inspectable end to end: the model, the quant, the kernel, the agent, and the wire. Two machines on one LAN are still a single trust boundary, and it belongs to the operator.
The record
35B on a laptop-class iGPU at interactive speed. An 8B that keeps up on an older board. Vision on the same engine. Six real bugs fixed and kept fixed. The harness, the numbers, and the methodology are all in the repo.