Card × model benchmark

Qwen2.5 Coder 7B on RTX 4070 Ti Q4_K_M

Smooth

On an RTX 4070 Ti (12 GB), Qwen2.5 Coder 7B in Q4_K_M generates 96,7 tokens/s and reads the prompt at 6 574 tokens/s, using 5,1 Go of video memory.

Median of 1 run · latest on 25/09/26

96,7 tok/s

Generation

6 574 tok/s

Prompt processing

5,1 Go

VRAM used (peak)

151 W

Average power

0,64 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 pageAll cards on Qwen2.5 Coder 7BFeel 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.

ContextKV cacheEstimated VRAMFits in 12 GB?Estimated speed
4 0960,2 Go5,3 Goyes≈ 93,1 tok/s
8 1920,4 Go5,5 Goyes≈ 89,1 tok/s
16 3840,9 Go5,9 Goyes≈ 82,0 tok/s
32 7681,8 Go6,8 Goyes≈ 70,8 tok/s

KV cache per token = 2 × 28 layers × 4 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”

Qwen2.5 Coder 7B on other cards

CardVRAMGenerationDifferencePrice
RTX 5070 Ti16 Go162 tok/s+67,6 %—
RTX 4070 Ti12 Go96,7 tok/sthis card—

Other models measured on the RTX 4070 Ti

Gemma 3 270M Q4_K_M 728 tok/sLlama 3.2 1B Q4_K_M 442 tok/sR1 Distill 1.5B Qwen Q4_K_M 328 tok/sGemma 2 2B Q4_K_M 202 tok/sLlama 3.2 3B Instruct Q4_K_M 189 tok/sGemma 3n E2B Q4_K_M 163 tok/sGemma 4 E2B Q4_K_M 198 tok/sGemma 3n E4B Q4_K_M 107 tok/sQwen2.5 7B Instruct Q4_K_M 98,8 tok/sR1 Distill 7B Qwen Q4_K_M 97,6 tok/sMeta Llama 3.1 8B Instruct Q4_K_M 91,9 tok/sR1 0528 (Qwen3) 8B Q4_K_M 90,1 tok/sGemma 4 E4B Q4_K_M 114 tok/sGemma 2 9B Q4_K_M 72,1 tok/sGemma 4 12B Q4_K_M 53,8 tok/sQwen2.5 14B Instruct Q4_K_M 50,5 tok/sDeepSeek R1 Distill Qwen 14B Q4_K_M 51,1 tok/sQwen2.5 Coder 14B Q4_K_M 50,3 tok/sPhi-4 Q4_K_M 50,8 tok/sCoder V2 Lite 16B-A2.4B Q4_K_M 189 tok/s

All measurements

DateSourceTypeGenerationPromptVRAMMax temp.Powerllama.cpp
25/09/26SephiGamecontext 32 768ASUS ROG Strix68,62 4627,1 Go68 °C200 Wllama.cpp 96278e39f
25/09/26SephiGamecontext 16 384ASUS ROG Strix80,74 1536,2 Go68 °C179 Wllama.cpp 96278e39f
25/09/26SephiGamecontext 8 192ASUS ROG Strix87,85 0875,7 Go66 °C172 Wllama.cpp 96278e39f
25/09/26SephiGamestandardASUS ROG Strix96,76 5745,1 Go62 °C151 Wllama.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