NVIDIA · workstation

RTX PRO 5000 Blackwell

RTX PRO 5000 Blackwell has 72 GB of VRAM at 1344 GB/s — about 66.96 GiB usable after driver and compositor overhead. 1982 of 2118 indexed models fit at 128K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
72 GB
GDDR7
Bandwidth
1344 GB/s
384-bit bus
Tensor FP16
295 TF
dense
TDP
300 W
$4569 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1695vision language 183audio asr 39image 2audio tts 21embedding 26video 16

What fits at 128K context

largest quantization that fits, per model · 1982 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
OYM-Qimi-122B-A10B-K2.6MoEIQ4_XS125B62.92 GiB3.00 GiB66.95 GiB0.01 GiB50±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEIQ4_XS123B62.92 GiB3.00 GiB66.95 GiB0.01 GiB50±37%
deepseek-coder-6.7B-kexerI1-IQ2_XS6.7B1.90 GiB64.00 GiB66.92 GiB0.04 GiB12±22%
Magicoder-S-DS-6.7BI1-IQ2_XS6.7B1.90 GiB64.00 GiB66.92 GiB0.04 GiB12±22%
deepseek-coder-6.7b-baseI1-IQ2_XS6.7B1.90 GiB64.00 GiB66.92 GiB0.04 GiB12±22%
Assistant_Pepe_70BQ2_K70.6B25.79 GiB40.00 GiB66.92 GiB0.04 GiB12±22%
WizardLM-7B-UncensoredI1-IQ2_XS6.7B1.90 GiB64.00 GiB66.92 GiB0.04 GiB12±22%
Llama-2-7B-32K-InstructI1-IQ2_XS6.7B1.90 GiB64.00 GiB66.92 GiB0.04 GiB12±22%
Luna-AI-Llama2-UncensoredI1-IQ2_XS6.7B1.90 GiB64.00 GiB66.92 GiB0.04 GiB12±22%
Swallow-7b-NVE-instruct-hfI1-IQ2_XS6.7B1.90 GiB64.00 GiB66.92 GiB0.04 GiB12±22%
Mistral-Small-4-119B-2603MoEQ4_0119B63.06 GiB2.81 GiB66.90 GiB0.06 GiB51±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ3_XXS109B41.87 GiB24.00 GiB66.89 GiB0.07 GiB15±37%
DeepSeek-V4-Flash-162BMoEKV unresolvedQ2_K92.2B55.09 GiB10.75 GiB66.89 GiB0.07 GiB27±37%
Hermes-4-70BUD-IQ3_XXS70.6B25.76 GiB40.00 GiB66.88 GiB0.08 GiB12±22%
Llama-3.3-70B-InstructUD-IQ3_XXS70.6B25.76 GiB40.00 GiB66.88 GiB0.08 GiB12±22%
DeepSeek-R1-Distill-Llama-70BUD-IQ3_XXS70.6B25.76 GiB40.00 GiB66.88 GiB0.08 GiB12±22%
SambaLingo-Japanese-ChatI1-IQ2_XXS6.9B1.83 GiB64.00 GiB66.85 GiB0.11 GiB12±22%
Mistral-Small-Instruct-2409Q3_K_S22.2B37.77 GiB28.00 GiB66.83 GiB0.13 GiB12±22%
CalmeRys-78B-Orpo-v0.1I1-IQ1_S78.0B22.62 GiB43.00 GiB66.75 GiB0.21 GiB12±22%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_XXS236B57.28 GiB8.44 GiB66.75 GiB0.21 GiB31±37%
DeepSeek-V2.5MoEIQ2_XXS236B57.28 GiB8.44 GiB66.75 GiB0.21 GiB31±37%
DeepSeek-Coder-V2-InstructMoEIQ2_XXS236B57.28 GiB8.44 GiB66.75 GiB0.21 GiB31±37%
c4ai-command-r-plus-08-2024IQ2_M104B33.56 GiB32.00 GiB66.74 GiB0.22 GiB12±22%
Apertus-70B-Instruct-2509IQ3_XXS70.6B25.53 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Maenad-70BI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Rombos-LLM-70b-Llama-3.3I1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
L3.3-Electra-R1-70bI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
L3.3-70B-Magnum-v4-SEIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Latxa-Llama-3.1-70B-Instruct-v2I1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Llama-3.3_70_b_uncensored_continuedI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Llama-3.3-70B-Instruct-abliteratedI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
grok-oss-Revenant-70BI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Llama-3.1-Nemotron-70B-Instruct-HFI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
L3.3-70B-Euryale-v2.3I1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Hermes-3-Llama-3.1-70BIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Hermes-4-70B-hereticI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Llama-3.1-70BIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Anubis-70B-v1.2IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Golem-70B-v1bI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
DeepSeek-R1-Distill-Llama-70B-hereticI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Legion-V2.1-LLaMa-70BI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Tess-R1-Limerick-Llama-3.1-70BIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
SEMIKONG-70BIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
functionary-medium-v3.2KV unresolvedIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Llama-3.1-WhiteRabbitNeo-2-70BIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
New-Dawn-Llama-3-70B-32K-v1.0I1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-IQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Athene-70BIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
L3.3-70B-Magnum-DiamondIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Meta-Llama-3-70B-InstructIQ3_XXS70.6B25.58 GiB40.00 GiB66.71 GiB0.25 GiB12±22%
Tess-3-Mistral-Nemo-12BF3212.2B45.63 GiB20.00 GiB66.68 GiB0.28 GiB12±22%
Lumimaid-v0.2-12BF3212.2B45.63 GiB20.00 GiB66.68 GiB0.28 GiB12±22%
MN-Violet-Lotus-12BF3212.2B45.63 GiB20.00 GiB66.68 GiB0.28 GiB12±22%
Mistral-Nemo-Instruct-2407F3212.2B45.63 GiB20.00 GiB66.68 GiB0.28 GiB12±22%
MN-12B-Celeste-V1.9F3212.2B45.63 GiB20.00 GiB66.68 GiB0.28 GiB12±22%
magnum-v2.5-12b-ktoF3212.2B45.63 GiB20.00 GiB66.68 GiB0.28 GiB12±22%
magnum-v2-12bF3212.2B45.63 GiB20.00 GiB66.68 GiB0.28 GiB12±22%
●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 RTX PRO 5000 Blackwell run?
1982 of 2118 indexed open-weight models fit a RTX PRO 5000 Blackwell at 131,072 context with f16 KV cache, the largest being OYM-Qimi-122B-A10B-K2.6 at IQ4_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a RTX PRO 5000 Blackwell actually have?
Its nameplate is 72 GB, but about 66.96 GiB is available to a model once driver and compositor overhead is accounted for.
Is a RTX PRO 5000 Blackwell fast for local AI?
Its memory bandwidth is 1344 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.