llamaperf

Qwen3.8 27B

on NVIDIA RTX 3090 · Ollama

Tone: mixedNVIDIA hardware
Nov 21, 2026
Throughput
43.5 t/s gen
Quant
Q4_K_M (GGUF)
VRAM reported
24 GB

Summary

User benchmarks Qwen3.8-27B on one RTX 3090 (24 GB) across Ollama, llama.cpp, LM Studio and vLLM. The imported UD-Q4_K_M GGUF in Ollama 0.40.15 ran at 43.5 tok/s on a short prompt without MTP; Ollama's qwen3.8:27b-mtp-q4_K_M library build reached 57.1 tok/s. Ollama used 17.0 GiB of GPU memory at 32k context, the least of the four, and handled the largest single-request context at 98,304 tokens. With MTP the largest context dropped to 61,440. At 32 concurrent requests Ollama produced 38.1 tok/s in total with the default single parallel slot, the lowest of the engines; OLLAMA_NUM_PARALLEL=4 raised it to 96.3, against 295.4 for llama.cpp with -np 32 and 345.8 for vLLM. The first answer after docker run took 88 s, 32.1 s of it model loading. Even with MTP it was 15 to 24% slower per request than LM Studio (65.8) and llama.cpp with MTP (70.9).

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

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 3090

Bars are medians on one scale, 0 to 111.7 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 3090 →Every Qwen3.8 report →Estimate Qwen3.8 on this card →