NVIDIA · consumer

GeForce RTX 2070 SUPER

GeForce RTX 2070 SUPER has 8 GB of VRAM at 448 GB/s — about 7.44 GiB usable after driver and compositor overhead. 919 of 2118 indexed models fit at 128K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
8 GB
GDDR6
Bandwidth
448 GB/s
256-bit bus
Tensor FP16
73 TF
dense
TDP
215 W
$499 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 742embedding 24vision language 86audio tts 20video 8audio asr 38image 1

What fits at 128K context

largest quantization that fits, per model · 919 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
OLMoE-1B-7B-0924-InstructMoEI1-IQ2_M6.9B2.17 GiB4.50 GiB7.44 GiB0.00 GiB37±37%
CycleGRPO-4BI1-IQ2_M4.8B1.57 GiB5.06 GiB7.44 GiB0.00 GiB48±12.9%
Jan-v3-4B-base-instructIQ2_M4.4B1.56 GiB5.06 GiB7.44 GiB0.00 GiB48±12.9%
Jan-code-4bIQ2_M4.4B1.56 GiB5.06 GiB7.44 GiB0.00 GiB48±12.9%
OmniAtlas-Qwen3-30B-A3BI1-IQ1_M31.7B6.59 GiB0.00 GiB7.44 GiB0.00 GiB48±12.9%
Qwen3-Omni-30B-A3B-CaptionerI1-IQ1_M31.7B6.59 GiB0.00 GiB7.44 GiB0.00 GiB48±12.9%
DeepSeek-Coder-V2-Lite-BaseMoEI1-IQ2_XS15.7B5.56 GiB1.07 GiB7.44 GiB0.00 GiB92±37%
DeepSeek-Coder-V2-Lite-InstructMoEIQ2_XS15.7B5.56 GiB1.07 GiB7.44 GiB0.00 GiB92±37%
DeepSeek-V2-Lite-ChatMoEIQ2_XS15.7B5.56 GiB1.07 GiB7.44 GiB0.00 GiB92±37%
EVA-Yi-1.5-9B-32K-V1I1-IQ3_XXS8.8B3.24 GiB3.38 GiB7.44 GiB0.00 GiB48±12.9%
Nemotron-3-Embed-8B-BF16IQ1_S8.0B1.81 GiB4.78 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-VL-4B-ThinkingUD-IQ3_XXS4.4B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-VL-4B-InstructUD-IQ3_XXS4.4B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-4BUD-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-4B-Thinking-2507UD-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Jan-nano-128kUD-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-4B-Instruct-2507UD-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Jan-nanoUD-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Gemma-4-12B-StyleTuneI1-IQ2_S13.0B4.20 GiB2.38 GiB7.43 GiB0.01 GiB48±12.9%
gemma-4-12b-heretic-styletune-headI1-IQ2_S12.0B4.20 GiB2.38 GiB7.43 GiB0.01 GiB48±12.9%
syrian-gemma-12bI1-IQ2_S13.0B4.20 GiB2.38 GiB7.43 GiB0.01 GiB48±12.9%
orpheus-3b-0.1-pretrainedQ5_13.8B2.68 GiB3.94 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-VL-4B-Instruct-Unredacted-MAXI1-IQ3_XXS4.4B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-VL-4B-Thinking-Unredacted-MAXI1-IQ3_XXS4.4B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Zubr1.0-VL-4BI1-IQ3_XXS4.4B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Huihui-Qwen3-VL-4B-Instruct-abliteratedI1-IQ3_XXS4.4B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-VL-4B-Instruct-UncensoredI1-IQ3_XXS4.4B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
OpenCaption-4B-VL-SFT-v1.0I1-IQ3_XXS4.4B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Parable-Qwen3-4B-Claude-Fable-5I1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Jan-v1-4BIQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-4b-Z-Image-Turbo-AbliteratedV1I1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-4B-abliteratedIQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Neuron-4B-InstructI1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
ChineseErrorCorrector4-4BI1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
FastContext-1.0-4B-SFT-abliteratedI1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-4B-Instruct_NSFW-V2.1I1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
FastContext-1.0-4B-SFTI1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
fable-traces-abliteratedI1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Nexa-AI-4B-InstructI1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Lumen-4B-InstructI1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-4B-Instruct-2507-hereticIQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-HereticLM-4BI1-IQ3_XXS4.0B1.56 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-VL-4B-Instruct-Uncensored-abliteratedQ2_K4.4B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
PopiT-Qwen3-4B-Medical-SFT-1128Q2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Logics-Parsing-v2Q2_K4.4B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Z-Image-Engineer-V6Q2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Huihui-Qwen3-4B-Instruct-2507-abliteratedQ2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Josiefied-Qwen3-4B-abliterated-v2Q2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-4B-abliterated-v2Q2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
CyberSecQwen-4BQ2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Huihui-Qwen3-4B-abliterated-v2Q2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-Reranker-4BQ2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Octen-Embedding-4BQ2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3-Embedding-4BQ2_K4.0B1.55 GiB5.06 GiB7.43 GiB0.01 GiB48±12.9%
Qwen3.5-9BQ4_K_M9.7B5.47 GiB1.13 GiB7.43 GiB0.01 GiB48±12.9%
Ministral-8B-Instruct-2410Q4_K_S8.0B4.36 GiB2.23 GiB7.43 GiB0.01 GiB48±12.9%
nomic-embed-codeQ5_K_S7.1B4.60 GiB1.97 GiB7.42 GiB0.02 GiB48±12.9%
SuperGemma-4-12b-abliteratedI1-Q2_K_S12.0B4.19 GiB2.38 GiB7.42 GiB0.02 GiB48±12.9%
gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-uncensored-hereticI1-Q2_K_S12.0B4.19 GiB2.38 GiB7.42 GiB0.02 GiB48±12.9%
gemma-4-12B-coder-fable5-composer2.5-v1-uncensored-hereticI1-Q2_K_S12.0B4.19 GiB2.38 GiB7.42 GiB0.02 GiB48±12.9%
●From the file◐Predictedwhat these mean

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Measured on this card

third-party benchmarks, aggregated
Workload◍MedianMiddle 50%Runs
Image generation7.02 it/s5.47–8.41287
Benchmarked· n=287

Aggregated from community-submitted runs, so the spread is wide by nature — it covers different models, resolutions, step counts and settings, not one controlled configuration. Read the middle 50% rather than the median alone. These figures are reproduced with attribution from vladmandic-sd-data-benchmark, which publishes no licence — so we display and link rather than redistribute them.

Questions people ask

What AI models can a GeForce RTX 2070 SUPER run?
919 of 2118 indexed open-weight models fit a GeForce RTX 2070 SUPER at 131,072 context with q4_0 KV cache, the largest being OLMoE-1B-7B-0924-Instruct at I1-IQ2_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a GeForce RTX 2070 SUPER actually have?
Its nameplate is 8 GB, but about 7.44 GiB is available to a model once driver and compositor overhead is accounted for.
Is a GeForce RTX 2070 SUPER fast for local AI?
Its memory bandwidth is 448 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.