GeForce RTX 5090 vs NVIDIA RTX Pro 6000 Blackwell 96GB for local LLMs

The NVIDIA RTX Pro 6000 Blackwell 96GB has more usable memory (96 GB vs 32 GB), so it runs 41 more of our tracked models at Q4, up to Mistral-Small-4-119B-2603. On our dated price snapshots the GeForce RTX 5090 is $6,566 cheaper. All checks are at Q4_K_M with about 4k context. Speed figures are estimates.

We may earn a commission when you buy through these links, at no extra cost to you.

What each one can run

Only the NVIDIA RTX Pro 6000 Blackwell 96GB runs: Mistral-Small-4-119B-2603, Youtu-LLM-A7B-Base, Laguna-S-2.1, gpt-oss-120b, c4ai-command-a-03-2025, GLM-4.5-Air, Llama-4-Scout-17B-16E-Instruct, Llama-4-Scout-17B-16E, GLM-4.5V, GLM-4.6V, Ling-2.6-flash, sarvam-105b, Ling-flash-2.0, LLaDA2.1-flash, Ring-flash-2.0, LLaDA2.2-flash, Solar-Open-100B, Qwen3-Next-80B-A3B-Instruct, Qwen3-Next-80B-A3B-Thinking, AliceAI-Foundation-80B-A3B-Base, Hunyuan-A13B-Instruct, Qwen3-Coder-Next, Kolibri-1, Qwen2.5-VL-72B-Instruct, UI-TARS-72B-DPO, Qwen2.5-72B-Instruct, Apertus-70B-Instruct-2509, Apertus-70B-2509, Llama-3.3-70B-Instruct, DeepSeek-R1-Distill-Llama-70B, Llama-3.1-70B, Hermes-2-Theta-Llama-3-70B, Meta-Llama-3.1-70B-Instruct, LongCat-Flash-Lite, AI21-Jamba2-Mini, Llama-3_1-Nemotron-51B-Instruct, Llama-3_3-Nemotron-Super-49B-v1_5, Llama-3_3-Nemotron-Super-49B-v1, Phi-3.5-MoE-instruct, GRIN-MoE, ALIA-40b.

Biggest at Q4: IQuest-Coder-V1-40B-Loop-Instruct on the GeForce RTX 5090, Mistral-Small-4-119B-2603 on the NVIDIA RTX Pro 6000 Blackwell 96GB. Memory decides what fits; it doesn't measure speed, which also depends on memory bandwidth and the runtime.

Estimated speed

Estimated decode tokens per second at Q4, from each card's memory bandwidth (how we estimate). MoE figures are upper bounds. Real speeds vary with the runtime, drivers and context length, and older cards (Pascal, Volta, Vega) may run below these estimates in some runtimes.

Model by model