Over the last few months I've been switching harnesses and engines until I finally landed on something good. I started with LM Studio + the Copilot extension and ended up here, so here are some tips for anyone running two 5060 Tis. Prefill is painful on these cards, so for hardware, make sure you have:

A motherboard with both GPUs on CPU lanes, either two slots wired to the CPU or one x16 slot with PCIe bifurcation (x8/x8). My second card used to hang off the chipset; moving it to CPU lanes cut time-to-first-token by about 2.5x and doubled my batching throughput.

A separate low-profile GPU just for system graphics. If one of your 5060 Tis drives the display, it burns 2-3 GB of VRAM, and since the layer split is even, the other card loses the same amount too. That's context you don't get.

Model: these are Blackwell cards, so an NVFP4 quant gives you better prefill and decode.

What I run: capy-flash (formerly ninfer-tp2) + OpenCode, Qwen3.8-27B NVFP4 with tensor parallel across both cards, MTP speculative decoding and int8 KV cache. I get 70-95 t/s decode and ~1,500 t/s prefill at 176K context (~3,500-4,000 t/s on shorter prompts). You can go to full context, I just don't have a third card.