Hardware and settings

RTX 3090 + RTX 3060 (omarchy, main node)

• GPUs: NVIDIA RTX 3090, 24 GB, and NVIDIA RTX 3060, 12 GB (36 GB of VRAM in all)
• CPU: Intel Core i9-11900K (8 cores, 16 threads)
• RAM: 80 GB
• OS: Arch Linux (Omarchy). Harnesses run in an incus container on the same host.

Each result names the GPU setup it ran on:

• RTX 3090 alone: the model runs on the 3090 only.
• RTX 3060 alone: the model runs on the 3060 only, the same budget as a 12 GB build.
• RTX 3090 + RTX 3060: the model's layers are split across both cards (llama.cpp layer split, or Strata with layers 0-11 on the 3060 and 12-47 on the 3090).
• RTX 3090 + RTX 3060 peer tier: Strata keeps the KV cache and dense weights on the 3090 and serves the experts from the 3060.

12GB 3060 (omarchy, earlier builds)

• The builds in the RTX 3060 section ran on this machine before the 3090 went in, on the RTX 3060 alone, with an Intel Core i5-10600 (6 cores, 12 threads) and 80 GB of RAM.

24GB M4 Mac Mini (earlier builds)

• Apple M4, 24 GB unified memory
• LM Studio (MLX and GGUF) for Mac-sized models
• The Mac mini results stay on the results page, but the machine left the lab on 2026-10-29, so no new builds come from it.

Orange Pi 5 Plus 16 GB (RK3588)

• Rockchip RK3588: 4x Cortex-A76 + 4x Cortex-A55, Mali-G610 GPU, 6 TOPS NPU
• 16 GB LPDDR4X, shared by the CPU, GPU and NPU
• Engines: rkllama on the NPU (RKLLM models), llama.cpp on Vulkan or the CPU (GGUF)
• Builds here use the Pi builder: a small step-by-step harness with checkpoint steers, sized for 2B to 9B models. Its engine speed results are listed on the results page.

Engines

• Strata: a mixture-of-experts engine that streams experts from RAM and disk, so a model bigger than the card's VRAM can run on the 3060, on the 3090, or on both together (Qwen3.8 Flash Next IQ3_XXS, Swift 1.5, the Coder).
• llama.cpp: dense and MoE GGUF models that fit (Bonsai 2 27B through PrismML's fork, Qwen3.6-35B-A3B, Ornith; on the 3090 also Saluki 27B, Sage A1.5, LFM 2.5 24B A2B and Qwen3.8 27B).
• LM Studio / MLX on the Mac mini (earlier builds).

Settings

Each model uses its card's recommended sampling preset. Thinking budget, context window and max output are recorded on every result page. The default for the pilot is medium thinking with a 3072-token budget and 7168 max output tokens.

Test machine

The checklist runs in headless Chromium on omarchy, pinned to two CPU cores and kept off the NVIDIA GPU, so it never competes with a model run.