Skip to content

What can the RTX 4070 Super run?

What fits on the RTX 4070 Super
Text models: 246Visual models: 12Strongest first
Memory
12 GB
Bandwidth
504 GB/s
Maker
NVIDIA

Open the calculator on this machineEvery model for 12 GB of VRAM

Text models that run on it

Text models that run on the RTX 4070 Super, best first
ModelBest fitContextSpeedStrongest at
Qwen3.5 9B8-bit24K37 tok/sChat, Coding, Agents, Docs
Gemma 4 12B6-bit112K34 tok/sChat, Coding, Docs
Qwen3.5 4B8-bit176K66 tok/sAgents, Docs
Ministral 3 14B4-bit16K42 tok/sCoding, Agents, Docs
Gemma 4 E4B8-bit128K59 tok/sChat, Coding, Docs
gpt-oss 20B3-bit40K156 tok/sAgents
Ministral 3 8B6-bit28K49 tok/sDocs
Gemma 4 E2B8-bit128K106 tok/sChat
Nemotron 3 Nano 4B8-bit256K82 tok/sDocs
Llama 3.1 8B8-bit20K42 tok/sChat, Coding, Agents, Docs
Qwen3 0.6B8-bit40K322 tok/sNot rated
Qwen3 14B3-bit24K49 tok/sNot rated
Qwen3 VL 8B Instruct6-bit24K50 tok/sText and images
Qwen3 8B8-bit16K42 tok/sNot rated
Qwen2.5 7B Instruct8-bit32K45 tok/sNot rated
Qwen2.5 VL 7B Instruct8-bit40K41 tok/sText and images
Unlimited OCR8-bit32K241 tok/sReading documents
Qwen2.5 1.5B Instruct8-bit32K182 tok/sNot rated
Qwen3.5 2B8-bit256K131 tok/sText and images
Qwen3 VL 4B Instruct8-bit40K69 tok/sText and images
Qwen3 1.7B8-bit40K144 tok/sNot rated
Qwen3 4B8-bit40K76 tok/sNot rated
Qwen3 4B Instruct 25078-bit44K76 tok/sNot rated
Qwen2.5 3B Instruct8-bit32K96 tok/sNot rated
DeepSeek OCR8-bit8K241 tok/sReading documents
Qwen2.5 Coder 14B Instruct3-bit20K48 tok/sCoding
Qwen2.5 VL 3B Instruct8-bit125K81 tok/sText and images
Qwen2.5 Coder 7B Instruct8-bit32K45 tok/sCoding
Mistral 7B Instruct V0.28-bit28K45 tok/sNot rated
Chandra OCR 28-bit160K59 tok/sReading documents
Qwen3 VL 2B Instruct8-bit72K138 tok/sText and images
Qwen2 VL 2B Instruct8-bit32K133 tok/sText and images
MiniCPM5 2B8-bit128K124 tok/sNot rated
Gemma 3 4B IT8-bit128K84 tok/sText and images
Llama 3.2 3B Instruct8-bit64K95 tok/sNot rated
LightOnOCR 2 1B8-bit16K258 tok/sText and images
Qwen2.5 14B Instruct3-bit20K48 tok/sNot rated
Qwen3 1.7B Base8-bit32K166 tok/sNot rated
HunyuanOCR8-bit96K244 tok/sText and images
Qwen3 0.6B Base8-bit32K379 tok/sNot rated
Rax 4.58-bit256K131 tok/sText and images
Qwen2 VL 7B Instruct8-bit32K41 tok/sText and images
Ornith 1.5 9B8-bit20K37 tok/sNot rated
Unlimited OCR 28-bit120K239 tok/sReading documents
Granite 4.1 3B8-bit88K87 tok/sNot rated
DeepSeek R1 Distill Qwen 1.5B8-bit128K182 tok/sNot rated
Mistral 7B V0.18-bit32K45 tok/sNot rated
Qwen2.5 Coder 3B Instruct8-bit32K96 tok/sCoding
Meta Llama 3 8B Instruct Abliterated V38-bit8K42 tok/sNot rated
EXAONE 3.5 7.8B Instruct8-bit20K43 tok/sNot rated
Saiga Llama3 8B8-bit8K42 tok/sNot rated
Qwen2.5 Coder 1.5B Instruct8-bit32K182 tok/sCoding
DeepSeek R1 0528 Qwen3 8B8-bit16K42 tok/sNot rated
Gemma 3 12B IT5-bit28K42 tok/sText and images
Phi 3.5 Mini Instruct8-bit16K82 tok/sNot rated
Qwen3 4B Thinking 25078-bit44K76 tok/sNot rated
Phi 4 Mini Instruct8-bit52K80 tok/sNot rated
Dots.ocr8-bit128K108 tok/sReading documents
DeepSeek R1 Distill Qwen 7B8-bit52K45 tok/sNot rated
Phi 3 Mini 4K Instruct8-bit4K82 tok/sNot rated

Best fit is the highest quality, up to 8-bit, that leaves the use case’s context. Ranked by our use-case scores, then by Hugging Face downloads. Speeds are rough, plus or minus 25%. The top 60 of 246 are shown.

Image and video models

Image and video models that run on the RTX 4070 Super
ModelBest fitPer outputStrongest at
Stable Diffusion XL 1.016-bit9.6 sImages
Z-Image Turbo8-bit13 sImages
FLUX.1 [dev]6-bit30 sImages
Qwen-Image 2.18-bit3.5 minImages
FLUX.3 Klein16-bit19 sImages
FLUX.2 [klein] 9B8-bit6.8 sImages
Krea 2 Turbo5-bit30 sImages
Stable Diffusion 3.5 Large8-bit1.8 minImages
MOSS-Image16-bit2.3 minImages
Wan 2.2 TI2V-5B16-bit20 minVideo
MOSS Image Base16-bit2.3 minImages
HunyuanVideo 1.56-bit1.7 hVideo

Only just too big

Questions

What is the best LLM for the RTX 4070 Super?

Of the open models we track, the strongest that run well on the RTX 4070 Super are Qwen3.5 9B, Gemma 4 12B, Qwen3.5 4B. 246 text models fit it in all.

Calculated from published specs, not benchmarked. All machines