By amount of VRAM
What can you run with 16 GB of VRAM? 27B models, big windows and the bandwidth trap
Sixteen gigabytes is the tier where the choice of card matters more than the amount of memory. Every 16 GB card loads the same models. They run at speeds that differ by a factor of three, and that gap is wider than the one between any two tiers below this. So this page has two jobs: what the memory buys, and how to avoid paying for memory you cannot use quickly.
What 16 GB buys
The usable memory is about 15.2 GB, and that lets a 24B to 27B dense model load at 4 bits, or a 12B to 14B at Q8 with a window to spare. These are the pairings we rank highest on an RTX 5060 Ti 16 GB.
| Pick | Model | Download | Memory | Speed |
|---|---|---|---|---|
| Best all-rounder | Qwen3.5 27B | IQ4_XS | 14.2 GB | 24 t/s |
| Best for code | Qwen2.5 Coder 14B Instruct | Q6_K | 12.6 GB | 27.1 t/s |
| Best for reasoning | Qwen3.5 27B | IQ4_XS | 14.2 GB | 24 t/s |
| Largest that still runs well | GLM 4.7 Flash | Q3_K_M | 14.8 GB | 92.7 t/s |
| Model | Parameters | Download | Memory | Speed | Longest context |
|---|---|---|---|---|---|
| Qwen3.5 27B | 27.78B | IQ4_XS | 14.2 GB | 24 t/s | 16k |
| Gemma 4 12B Instruct | 11.96B | Q8_0 | 13.6 GB | 25.1 t/s | 16k |
| GPT OSS 20B | 20.91B (4.18B active) | MXFP4 | 11.6 GB | 80.1 t/s | 128k |
| Qwen3 14B | 14.77B | Q6_K | 12.4 GB | 27.6 t/s | 16k |
| Gemma 3 27B Instruct | 27.43B | IQ4_XS | 15.1 GB | 22.5 t/s | 8k |
| Qwen3.5 9B | 9.65B | Q8_0 | 9.25 GB | 37.3 t/s | 128k |
| Qwen2.5 14B Instruct | 14.77B | Q6_K | 12.6 GB | 27.1 t/s | 16k |
| Qwen2.5 Coder 14B Instruct | 14.77B | Q6_K | 12.6 GB | 27.1 t/s | 16k |
| Gemma 4 26B A4B Instruct | 25.81B (3.74B active) | IQ4_XS | 14.0 GB | 71.6 t/s | 16k |
| Phi 4 | 14.66B | Q6_K | 12.8 GB | 26.7 t/s | 16k |
| Mistral Small 24B Instruct 2501 | 23.57B | Q4_K_M | 14.6 GB | 23.3 t/s | 8k |
| Gemma 3 12B Instruct | 12.19B | Q8_0 | 13.3 GB | 25.6 t/s | 32k |
We think two things stand out. The dense 27B models only fit at IQ4_XS and with a modest window, so you get the model and not much more. And the mixture-of-experts models are the real winners at this size: gpt-oss-20b loads whole, with several gigabytes left for context, and Gemma 4 26B A4B loads at 4 bits and generates at several times the speed of a dense 27B. If you care about speed, they are the models to start with. If you care about the best single answer, the dense 27B is stronger per token.
The trade at this tier is the same as the one between 8 and 12 GB, moved up a size. A 27B at 4 bits is a stronger model than a 14B at Q8, and the 14B leaves room for 32k or more of context. We lean towards the bigger model for chat and reasoning and the smaller one for anything that reads long files.
The bandwidth trap
The same models run at very different speeds across 16 GB cards, and the cheapest 16 GB card is the slowest by a wide margin. The table is Qwen3.5 27B at IQ4_XS and a 14B at Q6 on the common 16 GB options.
| Device | Memory | Bandwidth | Qwen3.5 27B | Qwen3 14B |
|---|---|---|---|---|
| GeForce RTX 4060 Ti 16GB | 16 GB | 288 GB/s | 15.4 t/s | 25.9 t/s |
| GeForce RTX 5060 Ti 16GB | 16 GB | 448 GB/s | 24.0 t/s | 40.3 t/s |
| Radeon RX 7600 XT | 16 GB | 288 GB/s | 13.1 t/s | 22.0 t/s |
| Radeon RX 9070 | 16 GB | 645 GB/s | 29.3 t/s | 49.2 t/s |
| GeForce RTX 4070 Ti SUPER | 16 GB | 672 GB/s | 35.9 t/s | 60.5 t/s |
| GeForce RTX 5070 Ti | 16 GB | 896 GB/s | 47.9 t/s | 80.7 t/s |
| GeForce RTX 5080 | 16 GB | 960 GB/s | 51.3 t/s | 86.4 t/s |
The RTX 4060 Ti 16 GB has 288 GB/s and an RTX 5070 Ti has 896, so the same 27B model runs at about 15.4 tokens per second on one and 47.9 on the other. At 15 tokens per second a 27B model is usable, but it is no fun, and at 45 it is. If you are choosing between a cheap 16 GB card and a more expensive one, the extra money buys speed, not capacity, and speed is what you will notice daily.
AMD's 16 GB cards, such as the RX 9070 and RX 7800 XT, have wide buses and strong value, and they work in llama.cpp and Ollama through ROCm or Vulkan. Intel's Arc A770 16 GB is the cheapest way into this tier on the used market. Check your exact card against the runtime you plan to use before buying, because support varies by card generation.
Which cards have 16 GB
We test the 16 GB tier by bandwidth first and brand second, since that is where the real differences are. The list is sorted that way, and the price column shows how little the fastest cards cost over the slowest.
| Device | Memory | Bandwidth | Price |
|---|---|---|---|
| GeForce RTX 5080 Desktop GPU | 16 GB | 960 GB/s | $1,050 |
| GeForce RTX 5070 Ti Desktop GPU | 16 GB | 896 GB/s | $780 |
| GeForce RTX 5080 Laptop Laptop GPU | 16 GB | 768 GB/s | n/a |
| GeForce RTX 4080 SUPER Desktop GPU | 16 GB | 736 GB/s | $950 |
| GeForce RTX 4080 Desktop GPU, used | 16 GB | 717 GB/s | $900 |
| GeForce RTX 4070 Ti SUPER Desktop GPU | 16 GB | 672 GB/s | $750 |
| Radeon RX 9070 XT Desktop GPU | 16 GB | 645 GB/s | $650 |
| Radeon RX 9070 Desktop GPU | 16 GB | 645 GB/s | $560 |
| Radeon RX 7800 XT Desktop GPU | 16 GB | 624 GB/s | $480 |
| GeForce RTX 4090 Laptop Laptop GPU | 16 GB | 576 GB/s | n/a |
| Radeon RX 7900 GRE Desktop GPU, used | 16 GB | 576 GB/s | $520 |
| Arc A770 16GB Desktop GPU, used | 16 GB | 560 GB/s | $280 |
| Radeon RX 6900 XT Desktop GPU, used | 16 GB | 512 GB/s | $400 |
| Radeon RX 6800 Desktop GPU, used | 16 GB | 512 GB/s | $330 |
The spread between the top and the bottom of the list is a factor of about 3, and it is the single most useful fact to carry into a purchase at this tier.
What 16 GB cannot do
It cannot hold a 32B model at 4 bits, which needs about 20.4 GB with a short window, so the 30B class belongs to the 24 GB tier. It cannot run a 70B model without most of it living in system RAM. And it leaves little room for the larger image and video pipelines, though SDXL and quantized Flux run comfortably. The 24 GB guide shows what the next step opens up.
Our advice at this tier
Buy for bandwidth. If the price gap between a narrow-bus 16 GB card and a wide one is under about 40%, take the wide one. Run a MoE model when you want speed and a dense 24B to 27B when you want the best answer, and keep the window at what you need, since every extra 8k comes out of the same memory as the models that are already close to the edge. And if you find yourself wanting the 30B models every day, the used RTX 3090 is the answer, which the GPU buying guide walks through.
Questions people ask
What is the best LLM for 16 GB of VRAM?
Qwen3.5 27B at IQ4_XS and Mistral Small 24B at Q4_K_M are the strongest dense models that load, and gpt-oss-20b fits whole with room for context. For a longer window, a 14B model at Q6 or Q8 is the better pick.
Is a 16 GB card enough for a 32B model?
Only at 3-bit precision, which costs quality, or with part of the model in system RAM. A 32B at Q4_K_M is about 20 GB, so it belongs on a 24 GB card.
Why is the RTX 4060 Ti 16 GB so slow?
Its 128-bit memory bus gives it 288 GB/s, less than half of what a 16 GB RTX 5070 Ti or an RX 9070 has. It holds the same models and generates them far more slowly.
Can I run a 70B model on 16 GB?
Not in any form we would recommend. At 4 bits a 70B model is around 42 GB, so most of it would sit in system RAM and generate at 1 to 3 tokens per second.