Qwen 27B vs Flash-Next on RTX 5060 Ti GPUs

01The goal

I saw a video about Strata as an optimized runner for Flash-Next and wondered whether it could improve my local AI setup—and maybe make it a bit faster. So I compared Qwen3.8-27B on llama.cpp (my previous setup) with Qwen3.8-Flash-Next on Strata.

02The workstation

Same Linux workstation throughout; 128K context, one active request. 27B: Q4_K_M/f16 KV. Flash-Next: IQ3_S/int8 KV.

Ryzen 9 9950X
96 GB RAM

GPU 0 · RTX 5060 Ti · 16 GB

PCIe 5 ×8
Fast slot

GPU 1 · RTX 5060 Ti · 16 GB

Borrowed from my personal PC

PCIe 3 ×1
Slow slot

CPU, RAM and GPU work together. Our one-GPU run used ~47 GiB of host memory, including reclaimable mapped weights. Strata 0.1.47 / release v0.2.2 · llama.cpp b11476.

Test 1 · Which model?Both GPUs active27B / llama.cpp vs Flash-Next / Strata

Test 2 · How many GPUs?Flash-Next: GPU 0 only vs bothFresh two-GPU control; same speed probes

03How I tested

Same prompts: 87 objective questions, mostly multiple-choice law quizzes, plus five fictional review tasks graded by one AI reviewer with model labels masked.

60 law quizzes24 workflows3 retrieval probes

Text only; sources inline, no attachments or web search. Quizzes used learned knowledge. Speed: three warm runs per input length, thinking off, exactly 128 output tokens, no prompt-cache reuse.

04What changed

With two GPUs, Flash-Next wrote 3.7–6.1× faster than 27B. Full 128-token replies were 1.5–2.4× faster. Four completed review tasks were 6.3–12.7× faster, with different answer lengths.

Flash-Next at 84, 6586 and 26087 input tokens: one GPU writes at 43.6, 64.3 and 61.4 tokens/s; two GPUs at 52.1, 74.1 and 72.9. Full replies take 3.8, 8.8 and 23.5 seconds on one GPU, versus 4.6, 15.5 and 33.2 on two. — Test 2 · Flash-Next, one vs two GPUs. Medians of three runs per input length; every answer capped at 128 tokens. Left/top: writing speed. Right/bottom: input processing plus writing. All three input lengths and raw GPU follow-up results.

One GPU reached 43.6 / 64.3 / 61.4 tokens/s across the three input lengths: 13–16% below two GPUs, but full replies arrived 18–43% sooner.

Original two-GPU quality	27B	Flash-Next
Law quiz · correct choice	47 / 60	51 / 60
Workflow content · audited	24 / 24	24 / 24
Five masked reviews · points	42 / 50	47 / 50

Law uses the separate answer-choice audit. Review points come from one AI reviewer. Workflow content treats [B] and B as the same source; original exact-field grades remain in the report.

These quizzes and synthetic reviews don’t establish real-work accuracy; public questions may have appeared in training.

05Conclusion

One fast-slot 5060 Ti was sufficient for these short replies. I’d use one card here. The second card’s slow bus may matter, but its effect wasn’t isolated. Next: longer reasoning answers, realistic document jobs, single-GPU accuracy and a faster second slot.