NVIDIA · consumer

GeForce RTX 5090

GeForce RTX 5090 has 32 GB of VRAM at 1792 GB/s — about 29.76 GiB usable after driver and compositor overhead. 2012 of 2118 indexed models fit at 8K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
32 GB
GDDR7
Bandwidth
1792 GB/s
512-bit bus
Tensor FP16
419 TF
dense
TDP
575 W
$1999 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1727audio tts 21vision language 181video 16image 2embedding 26audio asr 39

What fits at 8K context

largest quantization that fits, per model · 2012 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
phi-4F1614.7B27.31 GiB1.56 GiB29.73 GiB0.03 GiB44±12.9%
Phi-4-reasoningBF1614.7B27.31 GiB1.56 GiB29.73 GiB0.03 GiB44±12.9%
Phi-4-reasoning-plusBF1614.7B27.31 GiB1.56 GiB29.73 GiB0.03 GiB44±12.9%
Qwen3-TTS-12Hz-0.6B-BaseF32915M28.88 GiB0.00 GiB29.72 GiB0.04 GiB44±12.9%
Qwen2.5-7B-Instruct-1MF327.6B28.38 GiB0.44 GiB29.67 GiB0.09 GiB44±12.9%
DeepSeek-R1-Distill-Qwen-7BF327.6B28.38 GiB0.44 GiB29.67 GiB0.09 GiB44±12.9%
UI-TARS-7B-DPOF328.3B28.38 GiB0.44 GiB29.67 GiB0.09 GiB44±12.9%
Qwen2-7B-InstructF327.6B28.38 GiB0.44 GiB29.67 GiB0.09 GiB44±12.9%
Hercules-5.0-Qwen2-7BF327.6B28.38 GiB0.44 GiB29.67 GiB0.09 GiB44±12.9%
Kepler-8B-Instruct-v2F167.6B28.37 GiB0.44 GiB29.66 GiB0.10 GiB44±12.9%
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-IQ3_XXS79.7B28.68 GiB0.19 GiB29.66 GiB0.10 GiB251±37%
MiniCPM-o-2_6F328.7B28.37 GiB0.44 GiB29.66 GiB0.10 GiB44±12.9%
IQuest-Coder-V1-40B-InstructI1-Q5_K_M39.8B26.26 GiB2.50 GiB29.66 GiB0.10 GiB44±12.9%
GLM-4.5-Air-REAP-82B-A12BMoEIQ1_M81.9B27.38 GiB1.44 GiB29.64 GiB0.12 GiB130±37%
c4ai-command-r-plus-08-2024IQ2_XXS104B26.65 GiB2.00 GiB29.63 GiB0.13 GiB44±12.9%
NousCoder-14BBF1614.8B27.51 GiB1.25 GiB29.62 GiB0.14 GiB44±12.9%
Qwen3-14B-Claude-4.5-Opus-High-Reasoning-DistillBF1614.8B27.51 GiB1.25 GiB29.62 GiB0.14 GiB44±12.9%
Qwen3-14BBF1614.8B27.51 GiB1.25 GiB29.62 GiB0.14 GiB44±12.9%
Qwen3-14B-abliteratedBF1614.8B27.51 GiB1.25 GiB29.62 GiB0.14 GiB44±12.9%
Josiefied-Qwen3-14B-abliterated-v3BF1614.8B27.51 GiB1.25 GiB29.62 GiB0.14 GiB44±12.9%
Hermes-4-14BBF1614.8B27.51 GiB1.25 GiB29.62 GiB0.14 GiB44±12.9%
Qwen3-14B-BaseF1614.8B27.51 GiB1.25 GiB29.62 GiB0.14 GiB44±12.9%
Salience-1.5-ProMoEQ6_K_L36.0B28.66 GiB0.16 GiB29.62 GiB0.14 GiB229±37%
Qwable-v1MoEQ6_K_L36.0B28.66 GiB0.16 GiB29.62 GiB0.14 GiB229±37%
T-SearchMoEQ6_K_L36.0B28.66 GiB0.16 GiB29.62 GiB0.14 GiB229±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-TQ1_0109B27.25 GiB1.50 GiB29.58 GiB0.18 GiB142±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-IQ4_XS57.3B27.68 GiB1.05 GiB29.57 GiB0.19 GiB154±37%
Phi-3.5-MoE-instructMoEKV unresolvedQ5_K_L41.9B27.75 GiB1.00 GiB29.56 GiB0.20 GiB116±37%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.56 GiB0.20 GiB44±12.9%
Yi-34B-200K-DARE-megamerge-v8Q6_K34.4B26.78 GiB1.88 GiB29.54 GiB0.22 GiB44±12.9%
Nous-Hermes-2-Yi-34BI1-Q6_K34.4B26.78 GiB1.88 GiB29.54 GiB0.22 GiB44±12.9%
Nous-Capybara-limarpv3-34BI1-Q6_K34.4B26.78 GiB1.88 GiB29.54 GiB0.22 GiB44±12.9%
Apertus-70B-Instruct-2509UD-IQ3_XXS70.6B26.01 GiB2.50 GiB29.49 GiB0.27 GiB45±12.9%
Rombo-LLM-V3.0-Qwen-72bI1-IQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
Qwen2.5-72B-Instruct-abliteratedI1-IQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
HuatuoGPT-o1-72BIQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
MiroThinker-v1.0-72BI1-IQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
Malaysian-Qwen2.5-72B-InstructI1-IQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
Qwen2.5-72BI1-IQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
magnum-v4-72bI1-IQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
KAT-Dev-72B-ExpIQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
Homer-v1.0-Qwen2.5-72BIQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ2_S72.7B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
Qwen2.5-VL-72B-InstructIQ2_S73.4B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
UI-TARS-72B-DPOIQ2_S73.4B26.02 GiB2.50 GiB29.45 GiB0.31 GiB45±12.9%
archangel_sft-kto_llama30bIQ4_XS32.5B16.28 GiB12.19 GiB29.33 GiB0.43 GiB45±12.9%
Qwen3.5-88BMoEI1-Q2_K_S87.7B28.30 GiB0.19 GiB29.31 GiB0.45 GiB207±37%
Kimi-Linear-48B-A3B-InstructMoEQ4_K_L49.1B28.26 GiB0.24 GiB29.30 GiB0.46 GiB45±12.9%
Gemma-4-Novelist-Eclipse-31BQ6_K32.7B25.97 GiB2.42 GiB29.27 GiB0.49 GiB45±12.9%
Gemma-4-31B-StyleTuneQ6_K32.7B25.97 GiB2.42 GiB29.27 GiB0.49 GiB45±12.9%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ1_S139B26.55 GiB1.94 GiB29.27 GiB0.49 GiB139±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ1_S139B26.55 GiB1.94 GiB29.27 GiB0.49 GiB139±37%
Apriel-1.6-15b-ThinkerBF1614.9B26.88 GiB1.50 GiB29.23 GiB0.53 GiB45±12.9%
Assistant_Pepe_70BQ2_K70.6B25.79 GiB2.50 GiB29.22 GiB0.54 GiB45±12.9%
Wizard-Vicuna-30B-UncensoredI1-IQ4_XS32.5B16.15 GiB12.19 GiB29.21 GiB0.55 GiB45±12.9%
Hunyuan-A13B-InstructMoEQ2_K80.4B27.40 GiB1.00 GiB29.19 GiB0.57 GiB45±12.9%
ALIA-40b-fc-2606I1-Q5_K_M40.4B26.78 GiB1.50 GiB29.19 GiB0.57 GiB45±12.9%
ALIA-40b-instruct-2606I1-Q5_K_M40.4B26.78 GiB1.50 GiB29.19 GiB0.57 GiB45±12.9%
Qwen3-Coder-NextMoEUD-IQ3_S79.7B27.65 GiB0.75 GiB29.19 GiB0.57 GiB216±37%
●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 generation21.32 it/s11.90–34.75172
Prompt processing13493.29 tok/s10927.34–14983.7050
Text generation288.98 tok/s280.78–298.5234
Benchmarked· n=172

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 5090 run?
2012 of 2118 indexed open-weight models fit a GeForce RTX 5090 at 8,192 context with f16 KV cache, the largest being phi-4 at F16. That covers text, vision-language, image, video and speech models.
How much usable memory does a GeForce RTX 5090 actually have?
Its nameplate is 32 GB, but about 29.76 GiB is available to a model once driver and compositor overhead is accounted for.
Is a GeForce RTX 5090 fast for local AI?
Its memory bandwidth is 1792 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.