Phi-4-mini-reasoning

Phi family · 4.0B params

1recorded run

108.2best tok/s

1verified

Evidence: 1 sourced run · 1 independent source · first recorded Jul 12, 2026. What these labels mean.

What the Phi-4-mini-reasoning numbers mean

The fastest verified result on our board is 108.2 output tok/s on an RTX 3060 x2 at Q4_K_M with a 512-token context. That is about 9.2 ms per generated token — 500 tokens takes roughly 4.6 seconds. Prompt processing on the same record runs at 3875.2 tok/s, so a 2,000-token prompt is read in about 0.5 seconds. Peak VRAM was measured at 8.6 GB.

How much VRAM does Phi-4-mini-reasoning need?

Phi-4-mini-reasoning has 4.0B 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)	3.4 GB	8 GB
Q8_0 (8-bit)	5.7 GB	8 GB
FP16 (16-bit)	10.0 GB	12 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)

RTX 3060 x2108.2

Models with a similar footprint

Model	Params	Q4_K_M weights
gemma-3n-E4B-it-text	4B	2.2 GB
NVIDIA-Nemotron-3-Nano-4B	4B	2.2 GB
Qwen3-4B	4B	2.2 GB
Qwen3-4B-Thinking-2507	4B	2.2 GB
Qwen3.5-4B	4B	2.2 GB
Qwen3.5-4B-MTP	4B	2.2 GB

Ranked by distance from Phi-4-mini-reasoning's parameter count. Full benchmark board.

All benchmarks Methodology

GPU	Model	Quant	Backend	Ctx	tok/s	prompt tok/s	VRAM GB	Source
RTX 3060 x2 · 12 GB	Phi-4-mini-reasoning	Q4_K_M	llama.cpp 13f2b28	512	108.2 ✓	3875.2	8.6	src