100t/s on RTX 5090 24GB

Details

Running Qwen3.6-27B via vLLM on Olares One's RTX 5090 Laptop GPU 24GB achieved 85–85–100 t/s.

With MTP n=3 applied, it maintained 92–95% acceptance after warmup.
Max context was 75K, with a peak of 99.7 t/s.
On the same hardware, llama.cpp UD-Q4_K_XL/NVFP4 GGUF got 33–36 t/s, and vLLM v0.17 NVFP4 got 39 t/s.
The NVFP4 + MTP n=1 combo on vLLM 0.19.1 caused OOM due to a 2.37 GiB BF16 buffer for mtp.fc.
Switching to Lorbus/Qwen3.6-27B-int4-AutoRound, which keeps only MTP in BF16, pushed results up to 65 t/s at n=1 and 85–100 t/s at n=3.
The author summarized this combo as being about 3x faster than llama.cpp, and noted it's comparable to or higher than the 5090 32GB reference figure (78–80 t/s).

Key point