NVIDIA · consumer

GeForce RTX 4090 Laptop

GeForce RTX 4090 Laptop has 16 GB of VRAM at 576 GB/s — about 14.88 GiB usable after driver and compositor overhead. 1811 of 2118 indexed models fit at 64K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
16 GB
GDDR6
Bandwidth
576 GB/s
256-bit bus
Tensor FP16
—
dense
TDP
150 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1549vision language 159video 15image 2audio asr 39embedding 26audio tts 21

What fits at 64K context

largest quantization that fits, per model · 1811 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-UncensoredMoEI1-Q2_K33.6B11.53 GiB2.53 GiB14.88 GiB0.00 GiB46±37%
Salience-1.5-FlashMoEI1-IQ3_S31.1B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
Huihui-Qwen3-VL-30B-A3B-Instruct-abliteratedMoEI1-IQ3_S31.1B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-ABLITERATED-UNCENSOREDMoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
MiroThinker-v1.0-30BMoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
Qwen3-30B-A3B-YOYO-V5MoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
Qwen3-30B-A3B-Thinking-2507-Claude-4.5-Sonnet-High-Reasoning-DistillMoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
Huihui-Qwen3-30B-A3B-Thinking-2507-abliteratedMoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
Huihui-Qwen3-30B-A3B-Instruct-2507-abliteratedMoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
Qwen3-30B-A3B-abliterated-eroticMoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.87 GiB0.01 GiB72±37%
L3-DARKEST-PLANET-16.5BQ4_K_S16.5B9.03 GiB4.99 GiB14.87 GiB0.01 GiB30±12.9%
Huihui-Qwen3-Coder-30B-A3B-Instruct-abliteratedMoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.86 GiB0.02 GiB72±37%
Qwen3-Coder-30B-A3B-Instruct-RTPurboMoEI1-IQ3_S30.5B12.39 GiB1.69 GiB14.86 GiB0.02 GiB72±37%
OpenAI-gpt-oss-20B-Claude-4.5-Opus-Heretic-UncensoredMoEI1-Q4_K_S20.9B13.65 GiB0.43 GiB14.86 GiB0.02 GiB80±37%
gpt-oss-20b-uncensoredMoEI1-Q4_K_S20.9B13.65 GiB0.43 GiB14.86 GiB0.02 GiB80±37%
gpt-oss-safeguard-20bMoEI1-Q4_K_S21.5B13.65 GiB0.43 GiB14.86 GiB0.02 GiB80±37%
Huihui-gpt-oss-20b-BF16-abliterated-v2MoEI1-Q4_K_S20.9B13.65 GiB0.43 GiB14.86 GiB0.02 GiB80±37%
metatune-gpt20b-R1.09MoEI1-Q4_K_S21.5B13.65 GiB0.43 GiB14.86 GiB0.02 GiB80±37%
gpt-oss-20b-DerestrictedMoEQ4_K_S20.9B13.65 GiB0.43 GiB14.86 GiB0.02 GiB80±37%
Seed-OSS-36B-InstructUD-IQ2_XXS36.2B9.46 GiB4.50 GiB14.86 GiB0.02 GiB30±12.9%
Qwen3-Coder-30B-A3B-InstructMoEQ3_K_S30.5B12.38 GiB1.69 GiB14.86 GiB0.02 GiB72±37%
Qwen3-VL-30B-A3B-InstructMoEQ3_K_S31.1B12.38 GiB1.69 GiB14.86 GiB0.02 GiB72±37%
Qwen3-30B-A3B-abliteratedMoEQ3_K_S30.5B12.38 GiB1.69 GiB14.86 GiB0.02 GiB72±37%
Qwen3.8-27BQ3_K_M27.8B12.87 GiB1.13 GiB14.86 GiB0.02 GiB30±12.9%
Qwen3.6-27BQ3_K_M27.8B12.87 GiB1.13 GiB14.86 GiB0.02 GiB30±12.9%
Aurora-Code-1MoEI1-Q3_K_M34.7B13.70 GiB0.35 GiB14.86 GiB0.02 GiB137±37%
internlm2-math-plus-20bI1-Q4_K_S19.9B10.62 GiB3.38 GiB14.86 GiB0.02 GiB30±12.9%
Laguna-XS-2.1MoEIQ3_XXS33.4B13.30 GiB0.74 GiB14.84 GiB0.04 GiB115±37%
MN-GRAND-23.5B-Gutenberg-UNCENSORED-V2-GLM4.7-ThinkingI1-Q2_K23.4B8.29 GiB5.70 GiB14.84 GiB0.04 GiB30±12.9%
SOLAR-10.7B-Instruct-v1.0-uncensoredQ8_010.7B10.62 GiB3.38 GiB14.83 GiB0.05 GiB30±12.9%
Nous-Hermes-2-SOLAR-10.7BQ8_010.7B10.62 GiB3.38 GiB14.83 GiB0.05 GiB30±12.9%
SOLAR-10.7B-Instruct-v1.0Q8_010.7B10.62 GiB3.38 GiB14.83 GiB0.05 GiB30±12.9%
Qwen3-42B-A3B-2507-Thinking-Abliterated-uncensored-TOTAL-RECALL-v2-Medium-MASTER-CODERMoEI1-IQ2_XS42.4B11.68 GiB2.36 GiB14.83 GiB0.05 GiB62±37%
Skyfall-31B-v4.2IQ2_S31.4B10.10 GiB3.80 GiB14.82 GiB0.06 GiB30±12.9%
Trinity-2-Codestral-22B-v0.2Q3_K_M22.2B10.02 GiB3.94 GiB14.82 GiB0.06 GiB30±12.9%
Cydonia-v1.3-Magnum-v4-22BI1-Q3_K_M22.2B10.02 GiB3.94 GiB14.82 GiB0.06 GiB30±12.9%
Mistral-Small-22B-ArliAI-RPMax-v1.1I1-Q3_K_M22.2B10.02 GiB3.94 GiB14.82 GiB0.06 GiB30±12.9%
Mistral-Small-Drummer-22BQ3_K_M22.2B10.02 GiB3.94 GiB14.82 GiB0.06 GiB30±12.9%
magnum-v4-22bI1-Q3_K_M22.2B10.02 GiB3.94 GiB14.82 GiB0.06 GiB30±12.9%
Codestral-22B-v0.1Q3_K22.2B10.02 GiB3.94 GiB14.82 GiB0.06 GiB30±12.9%
Codestral-22B-v0.1-hfQ3_K_M22.2B10.02 GiB3.94 GiB14.82 GiB0.06 GiB30±12.9%
Mistral-MOE-4X7B-Dark-MultiVerse-Uncensored-Enhanced32-24BMoEQ3_K_L24.2B11.73 GiB2.25 GiB14.82 GiB0.06 GiB17±37%
dolphin-2.9.1-mixtral-1x22bMoEI1-Q3_K_M22.2B10.01 GiB3.94 GiB14.81 GiB0.07 GiB17±37%
Wan2.2-S2V-14BQ5_K_M16.3B13.97 GiB0.00 GiB14.81 GiB0.07 GiB30±12.9%
Qwen3.6-14B-A3B-FableVibesMoEQ8_013.8B13.65 GiB0.35 GiB14.80 GiB0.08 GiB97±37%
Qwen3.6-14B-A3B-VibeForged-v2MoEQ8_013.8B13.65 GiB0.35 GiB14.80 GiB0.08 GiB97±37%
gemma-4-26B-A4B-itMoEIQ4_XS26.5B13.23 GiB0.79 GiB14.80 GiB0.08 GiB30±12.9%
DA3-BASEF32—13.94 GiB0.00 GiB14.79 GiB0.09 GiB30±12.9%
gemma-4-A4B-98e-v6-coder-itMoEQ5_K_S20.5B13.21 GiB0.79 GiB14.78 GiB0.10 GiB30±12.9%
Nemotron-Mini-4B-InstructQ6_K4.2B11.72 GiB2.25 GiB14.78 GiB0.10 GiB30±12.9%
gemma-4-12B-coder-fable5-composer2.5-v1-abliteratedQ8_012.0B12.68 GiB1.26 GiB14.78 GiB0.10 GiB30±12.9%
gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5-abliteratedQ8_012.0B12.68 GiB1.26 GiB14.78 GiB0.10 GiB30±12.9%
dolphin-2.6-mixtral-8x7bMoEI1-IQ2_XXS46.7B11.69 GiB2.25 GiB14.78 GiB0.10 GiB41±37%
xLAM-8x7b-rMoEIQ2_XXS46.7B11.69 GiB2.25 GiB14.78 GiB0.10 GiB41±37%
Llama3.2-24B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-Q5_K_M18.0B12.00 GiB1.97 GiB14.78 GiB0.10 GiB56±37%
GRM-2.6-Plus-0628Q3_K_S27.8B12.78 GiB1.13 GiB14.77 GiB0.11 GiB30±12.9%
ThinkingCap-Qwen3.6-27BQ3_K_S27.4B12.78 GiB1.13 GiB14.77 GiB0.11 GiB30±12.9%
Tess-4-27BQ3_K_S27.8B12.78 GiB1.13 GiB14.77 GiB0.11 GiB30±12.9%
Qwen3-16B-A3BMoEQ6_K16.0B12.28 GiB1.69 GiB14.76 GiB0.12 GiB60±37%
Llama-3.2-3BF163.2B11.98 GiB1.97 GiB14.76 GiB0.12 GiB30±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 generation15.96 it/s10.58–21.15312
Benchmarked· n=312

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 4090 Laptop run?
1811 of 2118 indexed open-weight models fit a GeForce RTX 4090 Laptop at 65,536 context with q4_0 KV cache, the largest being Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-Uncensored at I1-Q2_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a GeForce RTX 4090 Laptop 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 4090 Laptop fast for local AI?
Its memory bandwidth is 576 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 4090 Laptop — what AI models can it run locally? — ossmodeldb