NVIDIA · consumer

GeForce RTX 3090

GeForce RTX 3090 has 24 GB of VRAM at 936 GB/s — about 22.32 GiB usable after driver and compositor overhead. 1360 of 2118 indexed models fit at 128K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
24 GB
GDDR6X
Bandwidth
936 GB/s
384-bit bus
Tensor FP16
142 TF
dense
TDP
350 W
$1499 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1106vision language 152embedding 26audio asr 38audio tts 21image 1video 16

What fits at 128K context

largest quantization that fits, per model · 1360 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4-32B-0414-Korean-CultureI1-IQ3_M32.6B13.80 GiB7.63 GiB22.32 GiB0.00 GiB31±12.9%
GLM-Z1-32B-0414IQ3_M32.6B13.80 GiB7.63 GiB22.32 GiB0.00 GiB31±12.9%
GLM-4-32B-0414IQ3_M32.6B13.80 GiB7.63 GiB22.32 GiB0.00 GiB31±12.9%
Salience-1.5-ProMoEQ4_036.0B19.01 GiB2.50 GiB22.32 GiB0.00 GiB87±37%
Qwable-v1MoEQ4_036.0B19.01 GiB2.50 GiB22.32 GiB0.00 GiB87±37%
T-SearchMoEQ4_036.0B19.01 GiB2.50 GiB22.32 GiB0.00 GiB87±37%
North-Mini-Code-1.0MoEQ4_130.5B17.97 GiB3.57 GiB22.31 GiB0.01 GiB65±37%
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16UD-Q4_K_S33.0B21.47 GiB0.00 GiB22.31 GiB0.01 GiB31±12.9%
Anubis-Mini-8B-v1Q5_K_M8.0B5.46 GiB16.00 GiB22.30 GiB0.02 GiB31±12.9%
Ministral-3-8B-Instruct-2512-BF16Q3_K_L8.9B4.45 GiB17.00 GiB22.29 GiB0.03 GiB31±12.9%
Luna-7B-A4BMoEIQ4_XS6.7B3.48 GiB18.00 GiB22.29 GiB0.03 GiB19±37%
Qwen3-Zero-Coder-Reasoning-V2-0.8BI1-Q4_1816M0.51 GiB21.00 GiB22.28 GiB0.04 GiB31±12.9%
Jan-code-4bQ6_K4.4B3.47 GiB18.00 GiB22.28 GiB0.04 GiB31±12.9%
legitus-instruct-v1I1-Q5_K_M8.1B5.41 GiB16.00 GiB22.28 GiB0.04 GiB31±12.9%
Apertus-8B-Instruct-2509I1-Q5_K_M8.1B5.41 GiB16.00 GiB22.28 GiB0.04 GiB31±12.9%
MiniCPM-Llama3-V-2_5BF168.5B5.43 GiB16.00 GiB22.27 GiB0.05 GiB31±12.9%
Qwen3.6-35B-A3B-Fable-5-DistillMoEI1-Q4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
Qwable-v2MoEI1-Q4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
Qwen3.6-35B-A3B-YOYO-V2MoEI1-Q4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
Ornith-1.0-35B-FP8-BLOCK-MTPMoEI1-Q4_K_S35.5B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
fable-coder-35B-A3BMoEI1-Q4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
Qwen3.6-35B-A3B-AntiLoopMoEI1-Q4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
PINQWEN-3.6-35B-CLEAN-BF16MoEI1-Q4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
UniMath-35B-A3BMoEI1-Q4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
Ornith-1.0-35B-Heretic-MTPMoEI1-Q4_K_S—18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
Fawen-1.0-35BMoEI1-Q4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
CyberStrike-OffSec-35BMoEQ4_K_S35.1B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-PreservedMoEI1-Q4_K_S35.1B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
Qwen3.6-35B-A3BMoEQ4_K_S36.0B18.97 GiB2.50 GiB22.27 GiB0.05 GiB87±37%
ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2I1-Q5_K_M21.8B14.45 GiB7.00 GiB22.27 GiB0.05 GiB31±12.9%
ERNIE-21B-A3B-Claude-4.5-High-OPUS-ThinkingI1-Q5_K_M21.8B14.45 GiB7.00 GiB22.27 GiB0.05 GiB31±12.9%
ERNIE-4.5-21B-A3B-ThinkingI1-Q5_K_M21.8B14.45 GiB7.00 GiB22.27 GiB0.05 GiB31±12.9%
Salience-1.5-FlashMoEI1-IQ2_M31.1B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
Huihui-Qwen3-VL-30B-A3B-Instruct-abliteratedMoEI1-IQ2_M31.1B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-ABLITERATED-UNCENSOREDMoEI1-IQ2_M30.5B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
MiroThinker-v1.0-30BMoEI1-IQ2_M30.5B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
Qwen3-30B-A3B-YOYO-V5MoEI1-IQ2_M30.5B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
Qwen3-30B-A3B-Thinking-2507-Claude-4.5-Sonnet-High-Reasoning-DistillMoEI1-IQ2_M30.5B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
Huihui-Qwen3-30B-A3B-Thinking-2507-abliteratedMoEI1-IQ2_M30.5B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
Huihui-Qwen3-30B-A3B-Instruct-2507-abliteratedMoEI1-IQ2_M30.5B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
Mistral-7B-v0.1KV unresolvedIQ4_XS7.2B5.43 GiB16.00 GiB22.26 GiB0.06 GiB31±12.9%
granite-8b-code-instruct-4kI1-IQ3_M8.1B3.43 GiB18.00 GiB22.26 GiB0.06 GiB31±12.9%
granite-8b-code-base-4kI1-IQ3_M8.1B3.43 GiB18.00 GiB22.26 GiB0.06 GiB31±12.9%
Huihui-Qwen3-Coder-30B-A3B-Instruct-abliteratedMoEI1-IQ2_M30.5B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
Qwen3-Coder-30B-A3B-Instruct-RTPurboMoEI1-IQ2_M30.5B9.47 GiB12.00 GiB22.26 GiB0.06 GiB29±37%
medgemma-27b-itI1-IQ3_XS28.8B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
gemma-3-27b-it-abliterated-refined-visionI1-IQ3_XS27.4B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
Nidum-Gemma-3-27B-it-UncensoredI1-IQ3_XS27.4B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
gemma-3-27b-it-abliteratedIQ3_XS27.4B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
AtomicGPT-gemma3-27bI1-IQ3_XS27.4B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
Unbound-v1.12.0-27BI1-IQ3_XS27.4B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
Mira-v1.12-Ties-27BI1-IQ3_XS27.4B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
gemma-3-27b-itIQ3_XS27.4B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
Medgamma27BI1-IQ3_XS27.0B10.77 GiB10.61 GiB22.26 GiB0.06 GiB31±12.9%
Nexa-AI-4x4B-InstructMoEI1-IQ2_XS12.1B3.45 GiB18.00 GiB22.26 GiB0.06 GiB19±37%
Olmo-3-7B-InstructIQ4_XS7.3B3.74 GiB17.69 GiB22.25 GiB0.07 GiB31±12.9%
Olmo-3-7B-ThinkIQ4_XS7.3B3.74 GiB17.69 GiB22.25 GiB0.07 GiB31±12.9%
Le-Chaton-Slim-23BMoEI1-IQ3_XXS23.3B8.44 GiB13.00 GiB22.25 GiB0.07 GiB25±37%
Aya-Medikal-V2I1-Q5_K_M8.0B5.40 GiB16.00 GiB22.25 GiB0.07 GiB31±12.9%
Ministral-3-8B-Instruct-2512-BF16-abliteratedIQ4_XS8.9B4.41 GiB17.00 GiB22.25 GiB0.07 GiB31±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 generation15.83 it/s12.16–19.572,819
Prompt processing4516.92 tok/s3384.60–5142.9029
Text generation153.01 tok/s121.53–162.4519
Benchmarked· n=2,819

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 3090 run?
1360 of 2118 indexed open-weight models fit a GeForce RTX 3090 at 131,072 context with f16 KV cache, the largest being GLM-4-32B-0414-Korean-Culture at I1-IQ3_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a GeForce RTX 3090 actually have?
Its nameplate is 24 GB, but about 22.32 GiB is available to a model once driver and compositor overhead is accounted for.
Is a GeForce RTX 3090 fast for local AI?
Its memory bandwidth is 936 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 3090 — what AI models can it run locally? — ossmodeldb