Qwen3-14B

Qwen family · 14.0B params

3recorded runs

90.8best tok/s

2verified

Evidence: 2 sourced runs · 1 awaiting review · 1 independent source · first recorded May 18, 2026 · last recorded Aug 11, 2026. What these labels mean.

What the Qwen3-14B numbers mean

The fastest verified result on our board is 90.8 output tok/s on an RTX PRO 6000 Blackwell at FP8 with a 32768-token context. That is about 11.0 ms per generated token — 500 tokens takes roughly 5.5 seconds. Across the 3 records we hold for this model the range is 35.8–90.8 tok/s on 3 GPUs, because quantisation, context length and the engine build all move the number.

How much VRAM does Qwen3-14B need?

Qwen3-14B has 14.0B parameters. These estimates use the same formula as the VRAM calculator: weights, plus about 15% for a moderate-context KV cache and runtime overhead, plus a small engine floor.

Quantisation	Estimated VRAM	Smallest card that fits
Q4_K_M (4-bit)	9.9 GB	12 GB
Q8_0 (8-bit)	17.9 GB	24 GB
FP16 (16-bit)	33.0 GB	48 GB

Long contexts and large batches raise the real number — read the smallest card as "possible with a reduced context". Measured throughput is on the benchmark board.

Fastest GPUs for this model (best tok/s)

Models with a similar footprint

Model	Params	Q4_K_M weights
InternVL3-14B	14B	7.9 GB
Ministral-3-14B-Reasoning-2512	14B	7.9 GB
DeepSeek-R1-Distill-Qwen-14B	15B	8.4 GB
Nemotron-Cascade-14B-Thinking	15B	8.4 GB
phi-4	15B	8.4 GB
Phi-4-reasoning	15B	8.4 GB

Ranked by distance from Qwen3-14B's parameter count. Full benchmark board.

All benchmarks Methodology

GPU	Model	Quant	Backend	Ctx	tok/s	prompt tok/s	VRAM GB	Source
RTX PRO 6000 Blackwell · 96 GB	Qwen3-14B	FP8	vllm 0.26.0	32768	90.8 ✓	—	—	src
RTX 3060 x2 · 12 GB	Qwen3-14B	Q4_K_M	llama.cpp 13f2b28	512	35.8 ✓	1136.8	15.4	src
RTX 3080 · 10 GB	Qwen3-14B	Q4_K_M	llama.cpp —	2048	75.0 …	—	8.7	src