NVIDIA · consumer

GeForce RTX 5070 Ti

GeForce RTX 5070 Ti has 16 GB of VRAM at 896 GB/s — about 14.88 GiB usable after driver and compositor overhead. 854 of 2118 indexed models fit at 128K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
16 GB
GDDR7
Bandwidth
896 GB/s
256-bit bus
Tensor FP16
176 TF
dense
TDP
300 W
$749 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 665vision language 101embedding 19video 15audio asr 35audio tts 18image 1

What fits at 128K context

largest quantization that fits, per model · 854 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Grug-12BIQ3_M12.0B5.56 GiB8.47 GiB14.87 GiB0.01 GiB45±12.9%
gemma-4-12B-it-Esper4IQ3_M12.0B5.56 GiB8.47 GiB14.87 GiB0.01 GiB45±12.9%
gemma-4-12B-itIQ3_M12.0B5.56 GiB8.47 GiB14.87 GiB0.01 GiB45±12.9%
Gemma-4-12B-StyleTuneI1-IQ3_S13.0B5.55 GiB8.47 GiB14.87 GiB0.01 GiB45±12.9%
gemma-4-12b-heretic-styletune-headI1-IQ3_S12.0B5.55 GiB8.47 GiB14.87 GiB0.01 GiB45±12.9%
syrian-gemma-12bI1-IQ3_S13.0B5.55 GiB8.47 GiB14.87 GiB0.01 GiB45±12.9%
Qwen3.6-28BMoEI1-Q3_K_S28.2B11.56 GiB2.50 GiB14.86 GiB0.02 GiB96±37%
Qwen3.5-28BMoEI1-Q3_K_S28.7B11.56 GiB2.50 GiB14.86 GiB0.02 GiB96±37%
nomic-embed-codeQ8_07.1B7.00 GiB7.00 GiB14.86 GiB0.02 GiB45±12.9%
EVA-Yi-1.5-9B-32K-V1I1-IQ1_M8.8B2.03 GiB12.00 GiB14.86 GiB0.02 GiB45±12.9%
Yi-Coder-9B-ChatIQ1_M8.8B2.03 GiB12.00 GiB14.86 GiB0.02 GiB45±12.9%
Ministral-3-3B-Instruct-2512UD-IQ2_XXS3.8B1.03 GiB13.00 GiB14.85 GiB0.03 GiB45±12.9%
Ministral-3-3B-Reasoning-2512UD-IQ2_XXS4.3B1.03 GiB13.00 GiB14.85 GiB0.03 GiB45±12.9%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ2_K_L20.9B11.03 GiB3.02 GiB14.84 GiB0.04 GiB71±37%
GLM-4.7-Flash-DerestrictedMoEI1-IQ2_XXS31.2B7.42 GiB6.61 GiB14.84 GiB0.04 GiB48±37%
Huihui-GLM-4.7-Flash-abliteratedMoEI1-IQ2_XXS31.2B7.42 GiB6.61 GiB14.84 GiB0.04 GiB48±37%
Felldude-Uncensored-Ministral3-3B-bf16I1-IQ2_XXS3.8B1.02 GiB13.00 GiB14.83 GiB0.05 GiB45±12.9%
Amaretto-3BI1-IQ2_XXS4.3B1.02 GiB13.00 GiB14.83 GiB0.05 GiB45±12.9%
Muse-Glimmer-30BUD-IQ3_XXS29.8B12.23 GiB1.72 GiB14.83 GiB0.05 GiB46±12.9%
Yi-1.5-6B-ChatQ8_06.1B6.00 GiB8.00 GiB14.82 GiB0.06 GiB45±12.9%
Wan2.2-S2V-14BQ5_K_M16.3B13.97 GiB0.00 GiB14.81 GiB0.07 GiB46±12.9%
Tiger-Gemma-12B-v3Q3_K_S12.8B5.49 GiB8.47 GiB14.80 GiB0.08 GiB46±12.9%
AfriqueGemma-12BI1-IQ3_S12.2B5.49 GiB8.47 GiB14.80 GiB0.08 GiB46±12.9%
diffusiongemma-26B-A4B-it-HERETIC-UncensoredMoETQ2_025.8B8.72 GiB5.29 GiB14.80 GiB0.08 GiB45±12.9%
North-Mini-Code-1.0MoEQ2_K_L30.5B10.45 GiB3.57 GiB14.79 GiB0.09 GiB74±37%
gemma-4-26B-A4BMoETQ2_026.5B8.71 GiB5.29 GiB14.79 GiB0.09 GiB45±12.9%
DA3-BASEF32—13.94 GiB0.00 GiB14.79 GiB0.09 GiB46±12.9%
Kimi-Linear-48B-A3B-InstructMoEIQ1_M49.1B10.17 GiB3.80 GiB14.78 GiB0.10 GiB46±12.9%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEI1-Q2_K_S23.0B7.36 GiB6.61 GiB14.78 GiB0.10 GiB47±37%
gpt-oss-20bMoEQ2_K_L21.5B10.95 GiB3.02 GiB14.76 GiB0.12 GiB72±37%
gpt-oss-safeguard-20bMoEQ2_K_L21.5B10.95 GiB3.02 GiB14.76 GiB0.12 GiB72±37%
ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2I1-Q2_K_S21.8B6.94 GiB7.00 GiB14.76 GiB0.12 GiB46±12.9%
ERNIE-21B-A3B-Claude-4.5-High-OPUS-ThinkingI1-Q2_K_S21.8B6.94 GiB7.00 GiB14.76 GiB0.12 GiB46±12.9%
ERNIE-4.5-21B-A3B-ThinkingI1-Q2_K_S21.8B6.94 GiB7.00 GiB14.76 GiB0.12 GiB46±12.9%
Ling-liteMoEIQ3_XXS16.8B6.96 GiB7.00 GiB14.75 GiB0.13 GiB45±37%
Goetia-26B-A4B-v1.3-Absolute-Heretic-ARAMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
Frank-26B-A4BMoEI1-IQ2_XXS26.5B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
G4-MeroMero-26B-A4B-it-uncensored-hereticMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
EVE-26b-XENO-HATMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
Gemma-4-26B-A4B-Animus-V14.1-FFT-hereticMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-qat-q4_0-unquantized-hereticMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
Huihui-gemma-4-26B-A4B-it-qat-q4_0-unquantized-abliteratedMoEI1-IQ2_XXS26.5B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
G4-MeroMero-26B-A4BMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
G4-Dark-Soul-26B-A4BMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-local-abliterated-sota-internal-t34MoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-SOMPOA-heresyMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-hereticMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-abliterixMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-heretic-ara-v2MoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
Gemma-4-26B-A4B-it-heretic-antislopMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-Claude-Opus-Distill-v2MoEI1-IQ2_XXS26.5B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-Heretic-StableMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-Uncensored-MAXMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-ultra-uncensored-hereticMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-ara-abliteratedMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
Huihui-gemma-4-26B-A4B-it-abliteratedMoEI1-IQ2_XXS26.5B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
Gemma-4-26B-A4B-AbliteratedMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma4-26b-fiction-bf16MoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
gemma-4-26B-A4B-it-heretic-araMoEI1-IQ2_XXS25.8B8.66 GiB5.29 GiB14.74 GiB0.14 GiB46±12.9%
granite-20b-code-instruct-8kQ5_K_L20.1B13.86 GiB0.00 GiB14.74 GiB0.14 GiB46±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 generation15.21 it/s9.27–17.7276
Prompt processing6291.54 tok/s5185.99–7492.5732
Text generation169.74 tok/s152.48–180.0020
Benchmarked· n=76

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 5070 Ti run?
854 of 2118 indexed open-weight models fit a GeForce RTX 5070 Ti at 131,072 context with f16 KV cache, the largest being Grug-12B at IQ3_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a GeForce RTX 5070 Ti actually have?
Its nameplate is 16 GB, but about 14.88 GiB is available to a model once driver and compositor overhead is accounted for.
Is a GeForce RTX 5070 Ti fast for local AI?
Its memory bandwidth is 896 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.
GeForce RTX 5070 Ti — what AI models can it run locally? — ossmodeldb