4 coding agents on one RTX 3090: ~290 tok/s total from a custom CUDA engine for Qwen3.8-27B (3-4x llama-server --parallel)
I've been building OpenJet, a local coding agent with its own CUDA megakernel engine for Qwen3.8-27B Q4_K_M. Until now it served one conversation at a time. The new release serves several agents at once on a single GPU.