NVIDIA · consumer

Titan Xp

Titan Xp has 12 GB of VRAM at 548 GB/s — about 11.16 GiB usable after driver and compositor overhead. 1461 of 2118 indexed models fit at 32K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
12 GB
GDDR5X
Bandwidth
548 GB/s
384-bit bus
Tensor FP16
—
dense
TDP
250 W
$1200 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
vision language 123text 1238video 14embedding 26audio tts 21image 1audio asr 38

What fits at 32K context

largest quantization that fits, per model · 1461 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiMo-VL-7B-RLI1-Q6_K8.3B5.83 GiB4.50 GiB11.16 GiB0.00 GiB38±12.9%
Kuwutu-7B-CYOA-v2I1-Q6_K7.6B5.83 GiB4.50 GiB11.16 GiB0.00 GiB38±12.9%
Wan2.1-FLF2V-14B-720PQ4_116.4B10.32 GiB0.00 GiB11.16 GiB0.00 GiB38±12.9%
GLM-4-32B-0414-Korean-CultureI1-IQ2_XXS32.6B8.36 GiB1.91 GiB11.15 GiB0.01 GiB38±12.9%
Wan2.1-I2V-14B-480PQ4_116.4B10.32 GiB0.00 GiB11.15 GiB0.01 GiB38±12.9%
Wan2.1-I2V-14B-720PQ4_116.4B10.32 GiB0.00 GiB11.15 GiB0.01 GiB38±12.9%
reka-flash-3.1I1-IQ1_S20.9B6.15 GiB4.13 GiB11.15 GiB0.01 GiB38±12.9%
zeta-2.1I1-Q6_K8.3B6.31 GiB4.00 GiB11.15 GiB0.01 GiB38±12.9%
gemma-4-12B-it-hereticQ5_K_M12.0B7.84 GiB2.47 GiB11.15 GiB0.01 GiB38±12.9%
Phi-3-medium-4k-instructI1-IQ2_S14.0B4.04 GiB6.25 GiB11.15 GiB0.01 GiB38±12.9%
Phi-3-medium-128k-instructIQ2_S14.0B4.04 GiB6.25 GiB11.15 GiB0.01 GiB38±12.9%
medgemma-27b-itI1-IQ2_XXS28.8B7.16 GiB3.11 GiB11.15 GiB0.01 GiB38±12.9%
gemma-3-27b-it-abliterated-refined-visionI1-IQ2_XXS27.4B7.16 GiB3.11 GiB11.15 GiB0.01 GiB38±12.9%
Nidum-Gemma-3-27B-it-UncensoredI1-IQ2_XXS27.4B7.16 GiB3.11 GiB11.15 GiB0.01 GiB38±12.9%
AtomicGPT-gemma3-27bI1-IQ2_XXS27.4B7.16 GiB3.11 GiB11.15 GiB0.01 GiB38±12.9%
Unbound-v1.12.0-27BI1-IQ2_XXS27.4B7.16 GiB3.11 GiB11.15 GiB0.01 GiB38±12.9%
Mira-v1.12-Ties-27BI1-IQ2_XXS27.4B7.16 GiB3.11 GiB11.15 GiB0.01 GiB38±12.9%
Medgamma27BI1-IQ2_XXS27.0B7.16 GiB3.11 GiB11.15 GiB0.01 GiB38±12.9%
Qwen3-15B-A2B-BaseMoEQ4_K_M15.6B8.84 GiB1.50 GiB11.15 GiB0.01 GiB84±37%
Grug-12BQ5_K_S12.0B7.83 GiB2.47 GiB11.14 GiB0.02 GiB38±12.9%
gemma-4-12B-it-Esper4Q5_K_S12.0B7.83 GiB2.47 GiB11.14 GiB0.02 GiB38±12.9%
gemma-4-12B-itQ5_K_S12.0B7.83 GiB2.47 GiB11.14 GiB0.02 GiB38±12.9%
MiroThinker-v1.0-8BQ5_K_L8.2B5.81 GiB4.50 GiB11.14 GiB0.02 GiB38±12.9%
Qwen3-8B-abliteratedQ5_K_L8.2B5.81 GiB4.50 GiB11.14 GiB0.02 GiB38±12.9%
Qwen3-8BQ5_K_L8.2B5.81 GiB4.50 GiB11.14 GiB0.02 GiB38±12.9%
Josiefied-Qwen3-8B-abliterated-v1Q5_K_L8.2B5.81 GiB4.50 GiB11.14 GiB0.02 GiB38±12.9%
Nemotron-Orchestrator-8BQ5_K_L8.2B5.81 GiB4.50 GiB11.14 GiB0.02 GiB38±12.9%
DeepSeek-R1-0528-Qwen3-8BQ5_K_L8.2B5.81 GiB4.50 GiB11.14 GiB0.02 GiB38±12.9%
Rocinante-XL-16B-v1I1-IQ1_S16.1B3.54 GiB6.75 GiB11.14 GiB0.02 GiB38±12.9%
Ministral-3-8B-Instruct-2512-BF16Q5_K_M8.9B6.04 GiB4.25 GiB11.13 GiB0.03 GiB38±12.9%
Kimi-VL-A3B-InstructMoEI1-Q4_116.4B9.37 GiB0.95 GiB11.13 GiB0.03 GiB91±37%
HomunculusQ2_K_L12.5B5.28 GiB5.00 GiB11.13 GiB0.03 GiB38±12.9%
Qwen3-VL-Embedding-8BQ6_K8.1B5.79 GiB4.50 GiB11.12 GiB0.04 GiB38±12.9%
Qwen3-8B-BaseQ6_K8.2B5.79 GiB4.50 GiB11.12 GiB0.04 GiB38±12.9%
qwen-indic-v1I1-Q6_K7.6B5.79 GiB4.50 GiB11.12 GiB0.04 GiB38±12.9%
Qwen3-Embedding-8BQ6_K7.6B5.79 GiB4.50 GiB11.12 GiB0.04 GiB38±12.9%
Snowpiercer-15B-v4-hereticI1-IQ2_XXS15.0B4.02 GiB6.25 GiB11.12 GiB0.04 GiB38±12.9%
Wan2.2-Distill-ModelsQ5_114.3B10.27 GiB0.00 GiB11.10 GiB0.06 GiB38±12.9%
Bernini-RQ5_114.3B10.26 GiB0.00 GiB11.10 GiB0.06 GiB38±12.9%
SkyReels-V2-DF-14B-540PQ5_114.3B10.27 GiB0.00 GiB11.10 GiB0.06 GiB38±12.9%
Maestro1-9BQ5_18.8B5.77 GiB4.50 GiB11.10 GiB0.06 GiB38±12.9%
Jan-v2-VL-highQ5_18.8B5.77 GiB4.50 GiB11.10 GiB0.06 GiB38±12.9%
Jan-v2-VL-medQ5_18.8B5.77 GiB4.50 GiB11.10 GiB0.06 GiB38±12.9%
MiniCPM-o-4_5Q5_19.4B5.77 GiB4.50 GiB11.10 GiB0.06 GiB38±12.9%
Fimbulvetr-11B-v2I1-IQ3_XS10.7B4.26 GiB6.00 GiB11.10 GiB0.06 GiB38±12.9%
Ling-liteMoEIQ4_XS16.8B8.55 GiB1.75 GiB11.10 GiB0.06 GiB72±37%
Devstral-Small-2-24B-Instruct-2512UD-IQ1_S24.0B5.18 GiB5.00 GiB11.10 GiB0.06 GiB38±12.9%
Mistral-Small-3.2-24B-Instruct-2506UD-IQ1_S24.0B5.18 GiB5.00 GiB11.09 GiB0.07 GiB38±12.9%
Devstral-Small-2507UD-IQ1_S23.6B5.18 GiB5.00 GiB11.09 GiB0.07 GiB38±12.9%
Devstral-Small-2505UD-IQ1_S23.6B5.18 GiB5.00 GiB11.09 GiB0.07 GiB38±12.9%
Magistral-Small-2507UD-IQ1_S23.6B5.18 GiB5.00 GiB11.09 GiB0.07 GiB38±12.9%
Mistral-Small-3.1-24B-Instruct-2503UD-IQ1_S24.0B5.18 GiB5.00 GiB11.09 GiB0.07 GiB38±12.9%
Apriel-1.6-15b-ThinkerI1-IQ2_XS14.9B4.25 GiB6.00 GiB11.09 GiB0.07 GiB38±12.9%
Qwen3.5-21B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingI1-IQ3_XS21.3B8.73 GiB1.50 GiB11.09 GiB0.07 GiB38±12.9%
Qwen3.6-21B-IQ-Ultra-Heretic-Uncensored-ThinkingI1-IQ3_XS21.3B8.73 GiB1.50 GiB11.09 GiB0.07 GiB38±12.9%
internlm2-math-plus-20bI1-IQ1_S19.9B4.23 GiB6.00 GiB11.09 GiB0.07 GiB38±12.9%
North-Mini-Code-1.0MoEUD-IQ2_M30.5B9.19 GiB1.13 GiB11.09 GiB0.07 GiB96±37%
granite-3.3-8b-instructQ5_K_S8.2B5.26 GiB5.00 GiB11.09 GiB0.07 GiB38±12.9%
granite-3.2-8b-instructQ5_K_S8.2B5.26 GiB5.00 GiB11.09 GiB0.07 GiB38±12.9%
UncensoredLM-DeepSeek-R1-Distill-Qwen-14BIQ2_S14.2B4.49 GiB5.75 GiB11.09 GiB0.07 GiB38±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.

Questions people ask

What AI models can a Titan Xp run?
1461 of 2118 indexed open-weight models fit a Titan Xp at 32,768 context with f16 KV cache, the largest being MiMo-VL-7B-RL at I1-Q6_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a Titan Xp actually have?
Its nameplate is 12 GB, but about 11.16 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Titan Xp fast for local AI?
Its memory bandwidth is 548 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.