Card × model benchmark

Qwen2.5 3B on RTX 5070 Ti Q4_K_M

Smooth

On an RTX 5070 Ti (16 GB), Qwen2.5 3B in Q4_K_M generates 305 tokens/s and reads the prompt at 14 781 tokens/s, using 3,3 Go of video memory.

Median of 1 run · latest on 25/09/26

305 tok/s

Generation

14 781 tok/s

Prompt processing

3,3 Go

VRAM used (peak)

103 W

Average power

2,96 tok/s/W

Efficiency

0,02 €/M tok

Electricity / M tokens · 0,25 €/kWh

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

RTX 5070 Ti pageAll cards on Qwen2.5 3BFeel 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 16 GB?Estimated speed
4 0960,1 Go3,5 Goyes≈ 287 tok/s
8 1920,3 Go3,6 Goyes≈ 269 tok/s
16 3840,6 Go3,9 Goyes≈ 238 tok/s
32 7681,1 Go4,5 Goyes≈ 194 tok/s

KV cache per token = 2 × 36 layers × 2 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 5070 Ti

Qwen2.5 Coder 0.5B Q4_K_M 805 tok/sQwen2.5 0.5B Q4_K_M 803 tok/sQwen3 0.6B Q4_K_M 728 tok/sQwen3.5 0.8B Q4_K_M 514 tok/sQwen2.5 1.5B Q4_K_M 481 tok/sQwen2.5 Coder 1.5B Q4_K_M 481 tok/sQwen3 1.7B Q4_K_M 456 tok/sQwen3.5 2B Q4_K_M 394 tok/sQwen2.5 Coder 3B Q4_K_M 305 tok/sLlama 3.2 3B Instruct Q4_K_M 283 tok/sQwen3 4B Q4_K_M 225 tok/sQwen3 2507 4B Q4_K_M 224 tok/sQwen3.5 4B Q4_K_M 196 tok/sQwen2.5 7B Instruct Q4_K_M 162 tok/sQwen2.5 Coder 7B Q4_K_M 162 tok/sMeta Llama 3.1 8B Instruct Q4_K_M 151 tok/sQwen3 8B Q4_K_M 146 tok/sQwen3.5 9B Q4_K_M 125 tok/sQwen2.5 Coder 14B Q4_K_M 81,9 tok/sQwen2.5 14B Instruct Q4_K_M 82,4 tok/sDeepSeek R1 Distill Qwen 14B Q4_K_M 81,5 tok/sQwen3 14B Q4_K_M 84,8 tok/sPhi-4 Q4_K_M 84,9 tok/s

All measurements

DateSourceTypeGenerationPromptVRAMMax temp.Powerllama.cpp
25/09/26French_Girl_13context 32 768MSI Gaming X Trio2046 3164,5 Go66 °C185 Wllama.cpp 96278e39f
25/09/26French_Girl_13context 16 384MSI Gaming X Trio2418 8203,9 Go62 °C151 Wllama.cpp 96278e39f
25/09/26French_Girl_13context 8 192MSI Gaming X Trio26410 8823,9 Go57 °C143 Wllama.cpp 96278e39f
25/09/26French_Girl_13standardMSI Gaming X Trio30514 7813,3 Go58 °C103 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