Card × model benchmark

Mistral 7B 7B on RTX 4070 Ti Q4_K_M

Smooth

On an RTX 4070 Ti (12 GB), Mistral 7B 7B in Q4_K_M generates 98,1 tokens/s and reads the prompt at 6 106 tokens/s, using 4,8 Go of video memory.

Median of 1 run · latest on 25/09/26

98,1 tok/s

Generation

6 106 tok/s

Prompt processing

4,8 Go

VRAM used (peak)

150 W

Average power

0,65 tok/s/W

Efficiency

0,11 €/M tok

Electricity / M tokens · 0,25 €/kWh

Estimate your monthly cost from your electricity price and usage: sign in

RTX 4070 Ti page All cards on Mistral 7B 7B Feel the speed Rent in the cloud

Where to buyLDLC Materiel.net TopAchat Cybertek Grosbill Rue du Commerce Amazon

What about a longer context?

An estimate, not a measurement: the KV cache grows with context, takes VRAM and slows generation down.

Context	KV cache	Estimated VRAM	Fits in 12 GB?	Estimated speed
4 096	0,5 Go	5,3 Go	yes	≈ 89,5 tok/s
8 192	1,0 Go	5,8 Go	yes	≈ 81,0 tok/s
16 384	2,0 Go	6,8 Go	yes	≈ 68,1 tok/s
32 768	4,0 Go	8,8 Go	yes	≈ 51,7 tok/s

KV cache per token = 2 × 32 layers × 8 KV heads × 128 × 2 bytes (f16). Estimated VRAM = measured VRAM + added KV cache. Speed ≈ measured speed × weights ÷ (weights + added KV cache) — attention compute is not modelled, the real drop is slightly larger. Fine-tune it in the “Configurator”

Other models measured on the RTX 4070 Ti

Gemma 3 270M Q4_K_M 728 tok/s LFM2.5 1.2B Q4_K_M 473 tok/s Llama 3.2 1B Q4_K_M 442 tok/s R1 Distill 1.5B Qwen Q4_K_M 328 tok/s Gemma 2 2B Q4_K_M 202 tok/s Llama 3.2 3B Instruct Q4_K_M 189 tok/s Ministral 3 3B Q4_K_M 173 tok/s Granite 4.1 3B Q4_K_M 166 tok/s Phi-4 mini 3.8B Q4_K_M 159 tok/s Gemma 3n E2B Q4_K_M 163 tok/s Gemma 4 E2B Q4_K_M 198 tok/s Gemma 3n E4B Q4_K_M 107 tok/s Qwen2.5 7B Instruct Q4_K_M 98,8 tok/s R1 Distill 7B Qwen Q4_K_M 97,6 tok/s Qwen2.5 Coder 7B Q4_K_M 96,7 tok/s Meta Llama 3.1 8B Instruct Q4_K_M 91,9 tok/s R1 0528 (Qwen3) 8B Q4_K_M 90,1 tok/s Ministral 3 8B Q4_K_M 84,0 tok/s Gemma 4 E4B Q4_K_M 114 tok/s Gemma 2 9B Q4_K_M 72,1 tok/s Gemma 4 12B Q4_K_M 53,8 tok/s Ministral 3 14B Q4_K_M 53,8 tok/s Qwen2.5 14B Instruct Q4_K_M 50,5 tok/s DeepSeek R1 Distill Qwen 14B Q4_K_M 51,1 tok/s Qwen2.5 Coder 14B Q4_K_M 50,3 tok/s Phi-4 Q4_K_M 50,8 tok/s Coder V2 Lite 16B-A2.4B Q4_K_M 189 tok/s

All measurements

Date	Source	Type	Generation	Prompt	VRAM	Max temp.	Power	llama.cpp
25/09/26	SephiGame	context 32 768ASUS ROG Strix	50,1	1 320	8,9 Go	67 °C	200 W	llama.cpp 96278e39f
25/09/26	SephiGame	context 16 384ASUS ROG Strix	67,6	2 458	6,9 Go	67 °C	187 W	llama.cpp 96278e39f
25/09/26	SephiGame	context 8 192ASUS ROG Strix	79,5	4 658	5,9 Go	68 °C	180 W	llama.cpp 96278e39f
25/09/26	SephiGame	standardASUS ROG Strix	98,1	6 106	4,8 Go	60 °C	150 W	llama.cpp 96278e39f

Protocol: llama-bench (llama.cpp), prompt 512 + generation 128 tokens, model fully loaded in VRAM. VRAM used = system peak (includes the OS, ~0.5 to 1 GB more than the model alone). Measure my card