VRAM tiers · measured gates across our fleet · Updated July 2026
What Can You Actually Run on 32GB of VRAM?
A 32GB GPU runs 10 of the 12 AI workloads we benchmark and is locked out of 2. 32GB is the first tier that runs modern image generation. FLUX.1-dev fits, FLUX.1 Kontext fits, Z-Image runs clean. What it still cannot do is hold a 70B language model or edit images with Qwen-Image-Edit, both of which want 42GB. Every floor below is measured: we ran each model on each card until it failed, and published where it failed.
NVIDIA RTX PRO 4500 Blackwell
32GB at 800 GB/s. AI Score 12.9/100 across our 12-workload suite. Measured on our bench.
- 32GB, runs 10 of our 12 workloads
- 800 GB/s of memory bandwidth
- Rentable by the hour
- Gated out of 2 of our 12 workloads
Best for: Testing whether 32GB is genuinely enough for your workload before you buy.
NVIDIA GeForce RTX 5090
32GB at 1792 GB/s. AI Score 22.3/100 across our 12-workload suite. Measured on our bench.
- 32GB, runs 10 of our 12 workloads
- 1792 GB/s of memory bandwidth
- Rentable by the hour
- Gated out of 2 of our 12 workloads
Best for: Testing whether 32GB is genuinely enough for your workload before you buy.
32GB is the first tier that runs modern image generation. FLUX.1-dev fits, FLUX.1 Kontext fits, Z-Image runs clean. What it still cannot do is hold a 70B language model or edit images with Qwen-Image-Edit, both of which want 42GB.
What runs on 32GB. Measured floors from our own gate testing
| Model | Type | VRAM floor | On 32GB? |
|---|---|---|---|
| Qwen3 4B | LLM | ~5GB | Runs |
| Llama 3.1 8B | LLM | ~8GB | Runs |
| Stable Diffusion XL | Image | ~8GB | Runs |
| Qwen2.5-Coder 14B | LLM | ~11.5GB | Runs |
| Z-Image Turbo | Image | ~13GB | Runs |
| LTX-Video | Video | ~14GB | Runs |
| Wan 2.2 (720p) | Video | ~18GB | Runs |
| Qwen3 32B | LLM | ~20GB | Runs |
| FLUX.1-dev | Image | ~26GB | Runs |
| FLUX.1 Kontext | Editing | ~26GB | Runs |
| Llama 3.3 70B | LLM | ~42GB | Won't fit (short 10GB) |
| Qwen-Image-Edit | Editing | ~42GB | Won't fit (short 10GB) |
Floors are measured at Q4_K_M for language models and BF16 for diffusion (SDXL at FP16). A smaller quantisation lowers the requirement and the quality, we don't mix precisions in one column.
What 32GB actually delivers, NVIDIA RTX PRO 4500 Blackwell, single stream, Q4_K_M
Measured on our bench. NVIDIA RTX PRO 4500 Blackwell is the highest-scoring 32GB card in our data (AI Score 12.9/100). Speed at a given tier varies with bandwidth, capacity decides what runs, bandwidth decides how fast.
Two things decide whether a GPU can do AI, and people consistently get the order wrong. Capacity decides what runs. Bandwidth decides how fast it runs. Capacity comes first, because a model that doesn't fit doesn't run slowly: it doesn't run at all, and no amount of bandwidth rescues it. That's why our AI Score treats a won't-fit as a zero rather than quietly excluding it. The practical consequence: shop for the VRAM tier that clears the models you actually intend to use, then optimise for bandwidth within that tier. Buying a faster card at the same capacity gets you a percentage. Buying the next capacity tier up can get you a model you literally could not run before.
Our verdict
32GB is the cheapest tier that runs modern diffusion. It's still 10GB short of a 70B. On our measured gates, 32GB clears 10 of 12 workloads and misses 2, the nearest being Llama 3.3 70B at ~42GB, short by 10GB. Buy capacity first and bandwidth second: capacity decides what runs, bandwidth only decides how fast.
FAQ
- Is 32GB of VRAM enough for AI?
- It depends entirely on which models, and the honest answer is a list rather than a yes or no. On our measured gates, 32GB runs 10 of our 12 workloads: Qwen3 4B, Llama 3.1 8B, Stable Diffusion XL, Qwen2.5-Coder 14B, Z-Image Turbo, LTX-Video and others. It is gated out of 2: Llama 3.3 70B, Qwen-Image-Edit. The nearest miss is Llama 3.3 70B, which needs about 42GB, short by 10GB.
- Can you run Llama 3.3 70B on 32GB?
- No. Llama 3.3 70B needs roughly 42GB at Q4_K_M and a 32GB card is 10GB short. This isn't a speed problem, the weights don't fit, so the model doesn't load. The largest LLM in our ladder that 32GB does run is Qwen3 32B.
- Can you run FLUX.1-dev on 32GB?
- Yes, FLUX.1-dev needs about 26GB at BF16 and 32GB clears it.
- What's the next VRAM tier worth stepping up to from 32GB?
- 42GB, and what it buys you is specifically Llama 3.3 70B, Qwen-Image-Edit. That's the useful way to think about a VRAM upgrade: not as a percentage more headroom, but as a named list of models you couldn't run before and can now. Stepping up a tier at the same bandwidth gets you capability. Buying a faster card at the same capacity gets you a percentage.
- Does a smaller quantisation get me past the VRAM limit?
- Often, yes, and we deliberately don't count it. Running a model at a lower quantisation reduces its memory requirement and its output quality, and if we silently swapped precision whenever a card ran out of room, every number on this site would become meaningless. Our floors are for Q4_K_M language models and BF16 diffusion. If you drop below that you're running a different model, and it should be measured as one.
- Do I need system RAM as well as 32GB of VRAM?
- For some workloads, yes, and it catches people out. LTX-Video wants roughly 20GB of host RAM and Wan 2.2 wants around 38GB, on top of their VRAM requirement, because the weights have to be staged. A card with enough VRAM sitting in a RAM-starved machine still fails. We gate on both, which is why a few cards you'd expect to clear a workload don't.
- I'm between two VRAM tiers and money's tight. Which way do I jump?
- Here's my rule after benchmarking the whole board: between 16GB and 24GB, go 16 and rent for the big jobs, the jump to 24 isn't big enough to matter once you account for context-window headroom on top of model weights, so they're effectively the same tier now. The real step up is 32GB on the newest architecture (RTX 5090). Below that, buy the cheaper card and put the savings into rented hours on cards you could never justify owning.
How we test
The VRAM floors on this page are measured, not calculated. We ran every model in our suite on every GPU in our fleet and recorded where it failed. When a model exceeds a card's VRAM we publish a hard won't-fit result with the requirement we observed, rather than quietly dropping to a smaller quantisation to produce a number, a card that can't run a model scores zero on it in our AI Score. LLMs are measured at Q4_K_M on llama.cpp (llama-bench, -p 512 -n 128). Diffusion and video run at BF16 on diffusers/ComfyUI, with SDXL at FP16. Floors are for those precisions: running a model at a smaller quantisation will lower its VRAM requirement and its quality, and we don't mix the two in one column. Two caveats worth stating plainly. First, these floors assume the model is the only thing on the card: a display attached to the same GPU, or a desktop compositor, eats into your headroom. Second, some workloads gate on system RAM as well as VRAM, LTX-Video wants roughly 20GB of host RAM and Wan 2.2 wants around 38GB, and a card with enough VRAM in a RAM-starved machine still fails. We gate on both. Speed figures are single-GPU, single-stream, batch-size-1, from our own runs. Cards labelled Measured were rented and run by us; Estimated cards are interpolated per workload against those anchors and labelled on every row.