What is the best LLM for 16GB of VRAM?

DeepSeek R1 Distill Qwen 14B (14.8B) is the largest model with comfortable headroom, at Q4_K_M, needing about 12GB against the ~15GB usable on a 16GB card. Mistral Codestral 22B (22.2B) is larger and will still load, but only as a tight fit with a short context. 12 of the 30 models in our directory fit at 16GB in some form.

Why is only about 15GB of a 16GB card usable?

The display output, the inference framework and the CUDA or Metal context all reserve memory before your model loads. Roughly 1GB is a realistic reservation, so a 16GB card gives you about 15GB to work with. Planning to the full 16GB is the most common reason a model that "should" fit fails to load.

Should I run a bigger model at Q4 or a smaller one at Q8?

Usually the bigger model at Q4_K_M. Above about 4 bits per weight, parameter count matters more than precision, so a larger model quantized harder generally beats a smaller model kept at full precision. Below Q4 that stops being true and quality falls away quickly.

How fast will these run on 16GB?

VRAM capacity decides what fits; memory bandwidth decides how fast it runs, and that varies by device. Among the 26 machines we profile at this capacity, bandwidth ranges from 102.4 to 960 GB/s — which can mean an order-of-magnitude difference in tokens per second for the same model. The hardware matrix gives an estimate per device.

Can I generate images with 16GB?

Yes — 10 open image models run at 1024×1024 on a card this size, the best being Stable Diffusion XL 1.0 at BF16 / FP16. Diffusion models load a large text encoder alongside the image model, so offloading that encoder to CPU matters more than it does for LLMs.

Best LLMs for 16GB VRAM

12 of 30 models fit on a 16GB card. The biggest with real headroom is DeepSeek R1 Distill Qwen 14B at Q4_K_M, though Mistral Codestral 22B (22.2B) will also load if you keep the context short. Below is every model scored against the ~15GB a 16GB card realistically gives you, with the best quantization each one can use.

8GB 12GB 16GB 24GB 32GB 48GB 64GB 128GB

Every model, scored for 16GB

Showing the best quantization each model can use at this capacity — higher precision first, dropping to Q4_K_M only where it has to.

Model	Best quant	Needs	Headroom	Fit
DeepSeek R1 Distill Qwen 14B14.8B · DeepSeek AI	Q4_K_M	12 GB	+3 GB	Comfortable
Qwen 2.5 Coder 14B Instruct14.7B · Alibaba Cloud	Q4_K_M	12 GB	+3 GB	Comfortable
Microsoft Phi-4 14B14.7B · Microsoft	Q4_K_M	12 GB	+3 GB	Comfortable
Mistral NeMo 12B Instruct12.2B · Mistral AI & NVIDIA	Q4_K_M	12 GB	+3 GB	Comfortable
Gemma 4 12B12B · Google DeepMind	Q4_K_M	10.5 GB	+4.5 GB	Comfortable
Google Gemma 2 9B9.24B · Google DeepMind	Q8_0	14 GB	+1 GB	Comfortable
Qwen3 8B8.2B · Alibaba Qwen	Q8_0	13.1 GB	+1.9 GB	Comfortable
Llama 3.1 8B Instruct8.03B · Meta AI	Q8_0	12 GB	+3 GB	Comfortable
DeepSeek R1 Distill Llama 8B8.0B · DeepSeek AI	Q8_0	12 GB	+3 GB	Comfortable
Qwen 2.5 Coder 7B Instruct7.6B · Alibaba Cloud	Q8_0	12 GB	+3 GB	Comfortable
Meta Llama 3.2 3B Instruct3.21B · Meta AI	BF16 / FP16	8 GB	+7 GB	Comfortable
Mistral Codestral 22B22.2B · Mistral AI	Q4_K_M	16 GB	—	Tight fit
DeepSeek V3 671B MoE671B (37B active) · DeepSeek AI	—	—	—	Will not fit
DeepSeek V4-Flash284B (13B active) · DeepSeek AI	—	—	—	Will not fit
OpenAI gpt-oss-120b116.8B (5.1B active) · OpenAI	—	—	—	Will not fit
Llama 4 Scout (17B active, 16 experts)109B (17B active) · Meta	—	—	—	Will not fit
Qwen 2.5 72B Instruct72.7B · Alibaba Cloud	—	—	—	Will not fit
DeepSeek R1 Distill Llama 70B70.6B · DeepSeek AI	—	—	—	Will not fit
Llama 3.3 70B Instruct70.6B · Meta AI	—	—	—	Will not fit
Qwen3.6 35B-A3B (MoE)36B (3B active) · Alibaba Qwen	—	—	—	Will not fit
Cohere Command R 35B35.0B · Cohere	—	—	—	Will not fit
DeepSeek R1 Distill Qwen 32B32.5B · DeepSeek AI	—	—	—	Will not fit
Qwen 2.5 Coder 32B Instruct32.5B · Alibaba Cloud	—	—	—	Will not fit
Gemma 4 31B31.3B · Google DeepMind	—	—	—	Will not fit
Qwen3-Coder 30B-A3B Instruct30.5B (3.3B active) · Alibaba Qwen	—	—	—	Will not fit
Qwen3.8 27B27.8B · Alibaba Qwen	—	—	—	Will not fit
Google Gemma 2 27B27.2B · Google DeepMind	—	—	—	Will not fit
Gemma 4 26B-A4B (MoE)26.5B (3.8B active) · Google DeepMind	—	—	—	Will not fit
Mistral Small 3 24B24.0B · Mistral AI	—	—	—	Will not fit
OpenAI gpt-oss-20b20.9B (3.6B active) · OpenAI	—	—	—	Will not fit

“Needs” is the recommended memory for that quantization, which leaves room for a working context. A 16GB card is treated as ~15GB usable, because the display, framework and CUDA or Metal context reserve roughly 1GB before your model loads.

Machines with about 16GB

We profile 26 machines at roughly this capacity. Their memory bandwidth ranges from 102.4 to 960 GB/s — and since generation speed is bandwidth-bound, two machines with the same 16GB can differ by an order of magnitude in tokens per second. Capacity decides what fits; bandwidth decides how fast.

High-End AI Rig (RTX 5080 16GB)960 GB/s · ~127.3 t/s est. (8B)Desktop with NVIDIA RTX 5070 Ti (16GB)896 GB/s · ~118.9 t/s est. (8B)Custom Workstation PC (RTX 4080 Super 16GB)736 GB/s · ~97.6 t/s est. (8B)Custom AI Workstation (RTX 4070 Ti Super 16GB)672 GB/s · ~89.1 t/s est. (8B)Desktop with AMD Radeon RX 9070 XT (16GB)640 GB/s · ~84.9 t/s est. (8B)Desktop PC (Radeon RX 7800 XT 16GB)624 GB/s · ~82.8 t/s est. (8B)Razer Blade 16 (RTX 4090 Mobile 16GB)576 GB/s · ~76.4 t/s est. (8B)Custom PC (Intel Arc A770 16GB)560 GB/s · ~74.3 t/s est. (8B)Desktop with NVIDIA RTX 5060 Ti (16GB)448 GB/s · ~59.4 t/s est. (8B)Desktop with AMD Radeon RX 9060 XT (16GB)320 GB/s · ~42.4 t/s est. (8B)

Image generation on 16GB

10 open image models run at 1024×1024 on a card this size — the best being Stable Diffusion XL 1.0 at BF16 / FP16, peaking near 9.2GB. Diffusion pipelines load a large text encoder alongside the image model, so offloading it matters more here than it does for language models.

Size a diffusion model for 16GB →

16GB VRAM: common questions

What is the best LLM for 16GB of VRAM?: DeepSeek R1 Distill Qwen 14B (14.8B) is the largest model with comfortable headroom, at Q4_K_M, needing about 12GB against the ~15GB usable on a 16GB card. Mistral Codestral 22B (22.2B) is larger and will still load, but only as a tight fit with a short context. 12 of the 30 models in our directory fit at 16GB in some form.
Why is only about 15GB of a 16GB card usable?: The display output, the inference framework and the CUDA or Metal context all reserve memory before your model loads. Roughly 1GB is a realistic reservation, so a 16GB card gives you about 15GB to work with. Planning to the full 16GB is the most common reason a model that "should" fit fails to load.
Should I run a bigger model at Q4 or a smaller one at Q8?: Usually the bigger model at Q4_K_M. Above about 4 bits per weight, parameter count matters more than precision, so a larger model quantized harder generally beats a smaller model kept at full precision. Below Q4 that stops being true and quality falls away quickly.
How fast will these run on 16GB?: VRAM capacity decides what fits; memory bandwidth decides how fast it runs, and that varies by device. Among the 26 machines we profile at this capacity, bandwidth ranges from 102.4 to 960 GB/s — which can mean an order-of-magnitude difference in tokens per second for the same model. The hardware matrix gives an estimate per device.
Can I generate images with 16GB?: Yes — 10 open image models run at 1024×1024 on a card this size, the best being Stable Diffusion XL 1.0 at BF16 / FP16. Diffusion models load a large text encoder alongside the image model, so offloading that encoder to CPU matters more than it does for LLMs.

Less memoryBest LLMs for 12GB VRAM →More memoryBest LLMs for 24GB VRAM →

Guides behind these numbers

How memory requirements are calculated, and which cards give you this much.

All VRAM tiers