Three Engines on One RTX 5090: TensorFold Runs Qwen3.8-27B at 268.2 Tok/s

NInfer leads at eight requests with 711.9 tok/s total; @MiaAI_lab’s one-click kit runs on 16 GB and 32 GB cards.

Sunday, October 25, 2026From 4 posts by @niklaslenz_ai

@niklaslenz_ai ran Qwen3.8-27B on one RTX 5090 with three engines, each set up the way its README says. All three got the same airship-dock prompt as an agent, one try, no help.

Engine	Code decode	8 at once, total	Build time	His rating
@BobClawblaw’s TensorFold 1.1.2 recipe + DFlash2	268.2 tok/s	204.6 tok/s	31.1 min	4/10
Neroued’s NInfer, NVFP4 + MTP3	169.0 tok/s	711.9 tok/s	56.5 min	5/10
@MiaAI_lab’s kit, EXL3 4.0 bpw + MTP	154.5 tok/s	None (one request at a time)	61.6 min	5/10

Source: @niklaslenz_ai. Swipe the table sideways to see every column.

His verdict: fastest for one request is TensorFold with DFlash2, fastest with eight at once is NInfer. Bob’s recipe serves one stream at full context, so eight requests share it and the total stays flat.

NInfer is a from-scratch C++/CUDA engine for exactly this card. At 240,000 tokens of context it used 28.9 GiB. For the eight-at-once number he ran it with 8 lanes; its command line says 2, and its README allows up to 8.

Mia’s kit installs with one click and runs on 16 GB and 32 GB cards. Between NInfer and the kit, he says, NInfer has more attention to detail, but Mia’s world has the better art style.

Three Engines on One RTX 5090: TensorFold Runs Qwen3.8-27B at 268.2 Tok/s

Sources