By amount of VRAM
What can you run with 32 GB of VRAM? The RTX 5090 tier
32 GB is a narrow tier. One consumer card has it, the RTX 5090, and a handful of workstation and older datacenter cards do too. What it buys is not a different class of model but a better version of the ones you could already run at 24 GB: higher precision, longer windows, and the fast mixture-of-experts models with room to breathe.
What fits
The usable memory is about 31.2 GB. A 27B to 32B model fits at Q6 or Q8, and a window of 16k to 64k stays affordable. These are the pairings the engine rates highest on an RTX 5090.
| Pick | Model | Download | Memory | Speed |
|---|---|---|---|---|
| Best all-rounder | Gemma 4 31B Instruct | Q6_K | 27.6 GB | 48.5 t/s |
| Best for code | Qwen3 32B | Q6_K | 26.9 GB | 49.9 t/s |
| Best for reasoning | Qwen3.5 27B | Q8_0 | 27.3 GB | 49.1 t/s |
| Largest that still runs well | Mixtral 8x7B Instruct V0.1 | Q4_K_M | 27.6 GB | 93 t/s |
| Model | Parameters | Download | Memory | Speed | Longest context |
|---|---|---|---|---|---|
| Gemma 4 31B Instruct | 31.27B | Q6_K | 27.6 GB | 48.5 t/s | 16k |
| Qwen3.5 27B | 27.78B | Q8_0 | 27.3 GB | 49.1 t/s | 64k |
| Qwen3 32B | 32.76B | Q6_K | 26.9 GB | 49.9 t/s | 16k |
| Granite 4.1 30B | 28.87B | Q8_0 | 31.0 GB | 43.1 t/s | 8k |
| Qwen2.5 32B Instruct | 32.76B | Q6_K | 26.9 GB | 49.9 t/s | 16k |
| Qwen2.5 Coder 32B Instruct | 32.76B | Q6_K | 26.9 GB | 49.9 t/s | 16k |
| Gemma 3 27B Instruct | 27.43B | Q8_0 | 28.7 GB | 46.7 t/s | 32k |
| Gemma 4 26B A4B Instruct | 25.81B (3.74B active) | Q8_0 | 26.7 GB | 169.1 t/s | 64k |
| Qwen3.5 35B A3B | 35.95B (2.9B active) | Q6_K | 27.6 GB | 318.7 t/s | 128k |
| GLM 4.7 Flash | 31.22B (3.66B active) | Q6_K | 24.4 GB | 240.5 t/s | 128k |
| Qwen3 30B A3B | 30.53B (3.34B active) | Q8_0 | 31.0 GB | 190.2 t/s | 8k |
| Mistral Small 24B Instruct 2501 | 23.57B | Q8_0 | 24.4 GB | 55.1 t/s | 32k |
Three things change from the 24 GB tier. The dense models go up a precision step: Qwen3.5 27B at Q8, which is as close to lossless as quantization gets, instead of Q6. The windows get longer, because the cache has somewhere to sit. And the mixture-of-experts models become remarkable. A 30B to 35B MoE model at Q6 or Q8 fits whole and generates at about 319 tokens per second, which is faster than you can read by an order of magnitude, and useful for agents that call a model many times.
The bandwidth deserves its own line. The RTX 5090 has 1792 GB/s, nearly twice the 3090, so a dense 32B at Q4 runs at about 66.0 tokens per second, where a 3090 does about 34.5. If a model fits both cards, this one is the fastest consumer option by a wide margin.
| Device | Memory | Bandwidth | Qwen3 32B | Qwen3.5 27B | Qwen3.5 35B A3B |
|---|---|---|---|---|---|
| GeForce RTX 3090 | 24 GB | 936 GB/s | 34.5 t/s | 44.3 t/s | 219 t/s |
| GeForce RTX 4090 | 24 GB | 1008 GB/s | 37.1 t/s | 47.7 t/s | 236 t/s |
| GeForce RTX 5090 | 32 GB | 1792 GB/s | 66.0 t/s | 84.8 t/s | 419 t/s |
What it cannot do
It cannot hold a 70B model. At 4 bits that is about 42.1 GB, so a quarter of it spills into system RAM and our engine estimates about 2.5 tokens per second, which is no longer interactive. The big mixture-of-experts models do better: gpt-oss-120b has a native file of 59 GiB, and with the experts split between the card and 64 GB of system RAM we estimate about 31.2 tokens per second. That is usable, and it needs the RAM. For a dense 70B that you want to chat with, 48 GB is the number to plan for, and our 48 GB guide shows how to get there.
A Mac with 32 GB is not the same thing
Apple Silicon machines with 32 GB of unified memory appear in tier lists next to the 5090 and are very different. macOS caps what the GPU may use at roughly 70% by default, so the effective figure is around 22 GB, and the bandwidth of a Max-class chip is under a third of the 5090's. They behave like a good 24 GB card at low speed, not like a 32 GB one. Our Mac and PC comparison explains where unified memory pays off, and it is at 64 GB and above.
Which devices have 32 GB
The list below mixes true 32 GB graphics cards with machines that have 32 GB of shared memory, and the two are not interchangeable. The unified-memory entries have a lot less bandwidth, and only part of their memory is available to the GPU.
| Device | Memory | Bandwidth | Price |
|---|---|---|---|
| GeForce RTX 5090 Desktop GPU | 32 GB | 1792 GB/s | $2,200 |
| Apple M1 Max 32GB Unified memory | 32 GB | 400 GB/s | $1,400 |
| Apple M2 Max 32GB Unified memory | 32 GB | 400 GB/s | $2,499 |
| Ryzen AI Max+ 395 32GB Unified memory | 32 GB | 256 GB/s | $1,299 |
| Jetson AGX Orin 32GB Unified memory | 32 GB | 204.8 GB/s | $1,799 |
| Apple M1 Pro 32GB Unified memory | 32 GB | 200 GB/s | n/a |
| Apple M2 Pro 32GB Unified memory | 32 GB | 200 GB/s | n/a |
| Apple M5 32GB Unified memory | 32 GB | 153 GB/s | $1,999 |
| Intel Core Ultra 9 288V 32GB Unified memory | 32 GB | 136 GB/s | n/a |
| Ryzen AI 9 HX 370 32GB Unified memory | 32 GB | 128 GB/s | n/a |
| Apple M4 32GB Unified memory | 32 GB | 120 GB/s | $999 |
If you see a 32 GB Mac or mini PC in this list and are tempted, read the next-lower tier first. Its real capacity is closer to 22 GB, and the 24 GB guide is the one that describes what it can run.
Practical notes on the RTX 5090
It is a 575 W card, so the power supply, the case and the 12V-2x6 cable all need attention, and it costs about 2,200 dollars at our last price check. Support for a new GPU architecture sometimes trails the hardware by weeks or months, so check your runtime's release notes before you buy on launch day. And because generation speed follows bandwidth, lowering the power limit by 20% costs only a few percent of speed and a good deal of heat, which is worth doing in a closed case.
Our advice at this tier
For local language models alone, the extra 8 GB over a used 3090 buys quality and context, not a new class of model, and it costs 3 times as much. Buy it if you also generate images and video, if you want the fastest possible answers from MoE models, or if the machine is for work and time is money. Otherwise 24 GB gets you most of the way, and two of them, as the multi-GPU guide explains, get you to 48.
Questions people ask
What can an RTX 5090 run that a 24 GB card cannot?
Mid-size dense models at higher quality and with more context. A 27B at Q8, a 31B at Q6 with a longer window, and larger mixture-of-experts models whole. The 70B class is still out of reach on one card.
Can a 32 GB card run a 70B model?
With heavy offload only. A 70B at 4 bits is about 42 GB, so roughly a quarter stays in system RAM and the speed drops to 2 or 3 tokens per second. Two 24 GB cards are a better way to run one.
Is a Mac with 32 GB the same as a 32 GB graphics card?
No. macOS lets the GPU use only about 70% of the memory by default, so a 32 GB Mac behaves like a 22 GB card, with far lower bandwidth than an RTX 5090.
Is 32 GB worth the price over 24 GB?
For local LLMs alone, usually not. The step from 24 to 32 GB buys higher quantization and longer windows on models you could already run, and the card costs much more than a used 24 GB one. It makes sense if you also generate images and video or want the fastest card there is.