Is 24GB 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, 24GB runs 8 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 4: FLUX.1-dev, FLUX.1 Kontext, Llama 3.3 70B, Qwen-Image-Edit. The nearest miss is FLUX.1-dev, which needs about 26GB, short by 2GB.

Can you run Llama 3.3 70B on 24GB?

No. Llama 3.3 70B needs roughly 42GB at Q4_K_M and a 24GB card is 18GB 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 24GB does run is Qwen3 32B.

Can you run FLUX.1-dev on 24GB?

No. FLUX.1-dev needs roughly 26GB at BF16 and a 24GB card is 2GB short. This is the cliff that catches most people out, because 24GB handles SDXL without complaint and feels like plenty right up until you try to load a modern image model. Running FLUX at a smaller quantisation is possible but it isn't the same model, which is why we don't publish it in the same column.

What's the next VRAM tier worth stepping up to from 24GB?

26GB, and what it buys you is specifically FLUX.1-dev, FLUX.1 Kontext. 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 24GB 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.

VRAM tiers · measured gates across our fleet · Updated July 2026

What Can You Actually Run on 24GB of VRAM?

A 24GB GPU runs 8 of the 12 AI workloads we benchmark and is locked out of 4. 24GB is the tier everyone assumes is enough, and it's the single most consequential misunderstanding in local AI. It runs Qwen3 32B comfortably. It runs SDXL and Z-Image with room. And it misses FLUX.1-dev, the defining image model of this generation, by two gigabytes. Every floor below is measured: we ran each model on each card until it failed, and published where it failed.

Rentable at this tier

NVIDIA GeForce RTX 4090

24GB at 1008 GB/s. AI Score 10.4/100 across our 12-workload suite. Measured on our bench.

Pros

24GB, runs 8 of our 12 workloads
1008 GB/s of memory bandwidth
Rentable by the hour

Cons

Gated out of 4 of our 12 workloads

Best for: Testing whether 24GB is genuinely enough for your workload before you buy.

Rentable at this tier

NVIDIA RTX PRO 4000 Blackwell

24GB at 672 GB/s. AI Score 7.6/100 across our 12-workload suite. Measured on our bench.

Pros

24GB, runs 8 of our 12 workloads
672 GB/s of memory bandwidth
Rentable by the hour

Cons

Gated out of 4 of our 12 workloads

Best for: Testing whether 24GB is genuinely enough for your workload before you buy.

Rentable at this tier

NVIDIA L4

24GB at 300 GB/s. AI Score 5.0/100 across our 12-workload suite. Measured on our bench.

Pros

24GB, runs 8 of our 12 workloads
300 GB/s of memory bandwidth
Rentable by the hour

Cons

Gated out of 4 of our 12 workloads

Best for: Testing whether 24GB is genuinely enough for your workload before you buy.

8/12

Workloads that run

at the precision we test

4/12

Workloads gated out

hard VRAM ceiling, no fix

26GB

Next thing you'd unlock

FLUX.1-dev, FLUX.1 Kontext

24GB cards in our data

of the fleet we've benchmarked

24GB is the tier everyone assumes is enough, and it's the single most consequential misunderstanding in local AI. It runs Qwen3 32B comfortably. It runs SDXL and Z-Image with room. And it misses FLUX.1-dev, the defining image model of this generation, by two gigabytes.

What runs on 24GB. Measured floors from our own gate testing

Model	Type	VRAM floor	On 24GB?
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	Won't fit (short 2GB)
FLUX.1 Kontext	Editing	~26GB	Won't fit (short 2GB)
Llama 3.3 70B	LLM	~42GB	Won't fit (short 18GB)
Qwen-Image-Edit	Editing	~42GB	Won't fit (short 18GB)

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 24GB actually delivers, NVIDIA GeForce RTX 4090, single stream, Q4_K_M

Qwen3 4B

260.54 tok/s

Llama 3.1 8B

171.29 tok/s

Qwen2.5-Coder 14B

95.12 tok/s

Qwen3 32B

44.28 tok/s

Measured on our bench. NVIDIA GeForce RTX 4090 is the highest-scoring 24GB card in our data (AI Score 10.4/100). 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

24GB clears 32B language models and misses FLUX.1-dev by 2GB. That two-gigabyte gap is the most expensive rounding error in AI hardware. On our measured gates, 24GB clears 8 of 12 workloads and misses 4, the nearest being FLUX.1-dev at ~26GB, short by 2GB. Buy capacity first and bandwidth second: capacity decides what runs, bandwidth only decides how fast.

FAQ

Is 24GB 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, 24GB runs 8 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 4: FLUX.1-dev, FLUX.1 Kontext, Llama 3.3 70B, Qwen-Image-Edit. The nearest miss is FLUX.1-dev, which needs about 26GB, short by 2GB.
Can you run Llama 3.3 70B on 24GB?: No. Llama 3.3 70B needs roughly 42GB at Q4_K_M and a 24GB card is 18GB 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 24GB does run is Qwen3 32B.
Can you run FLUX.1-dev on 24GB?: No. FLUX.1-dev needs roughly 26GB at BF16 and a 24GB card is 2GB short. This is the cliff that catches most people out, because 24GB handles SDXL without complaint and feels like plenty right up until you try to load a modern image model. Running FLUX at a smaller quantisation is possible but it isn't the same model, which is why we don't publish it in the same column.
What's the next VRAM tier worth stepping up to from 24GB?: 26GB, and what it buys you is specifically FLUX.1-dev, FLUX.1 Kontext. 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 24GB 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.