Qwen3.8 27B
on 2× NVIDIA RTX 5060 Ti 16GB · SGLang
Use cases
math
Summary
User reports Qwen3.8-27B at approximately 126 tok/s single-stream decode on two RTX 5060 Ti 16 GB cards.
Setup is SGLang 0.5.23 with NVFP4 weights, FP8 KV cache, tensor parallelism TP=2, and a DFlash2 drafter with 8 draft tokens; the KV pool holds 126,585 tokens against a 131,072 advertised context.
Decode falls to 107.4 tok/s at 60K context (2,681 tok/s prefill) and 87.3 tok/s at 124K (1,577 tok/s prefill); a rerun at 126,000 context gave 124 and 85 tok/s. Across about a dozen DFlash configurations the short-prompt median stayed between 120 and 128 tok/s, and the two cards draw roughly 220-270 W during decode.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
Qwen3.8 27B at 4-bit by hardware
- NVIDIA RTX 509032 GB80.0 t/s · 7 runs
- 67.0 t/s · one run
- NVIDIA RTX 5070 Ti16 GB54.7 t/s · one run
- 50.0 t/s · one run
- NVIDIA RTX 409024 GB40.0 t/s · 3 runs
- AMD RX 7900 XTX24 GB40.0 t/s · 7 runs
Bars are medians on one scale, 0 to 94.0 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.
Models most run on the NVIDIA RTX 5060 Ti 16GB
- Qwen3.8 27B3-bit28.8 t/s · 2 runs
- 95.2 t/s · one run
- Gemma 4 8B4-bit73.9 t/s · one run
- Mimo 2.6 9B8-bit41.6 t/s · one run
- Muse 30B4-bit18.0 t/s · one run
- Qwen2.5 7B4-bit77.0 t/s · one run
Bars are medians on one scale, 0 to 95.2 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.
Everything reported on the NVIDIA RTX 5060 Ti 16GB →Every Qwen3.8 report →Estimate Qwen3.8 on this card →