RunsOnMy ModelsGPUs & MacsBy VRAMAPI pricesVRAM calculatorNew

AI models that fit in 12 GB of VRAM

154 of 285 open models we track fit at 4-bit (Q4_K_M) with an 8K context on a 12 GB discrete GPU. About 11.4 GB is usable. 136 of those also fit at 8-bit. Typical 12 GB cards: RTX 3060 12GB, RTX 4070, RTX 5070 Ti 12GB.

154models fit at 4-bit, 8K
136also fit at 8-bit
16Blargest that fits
11.4 GBusable of 12 GB

GPUs and Macs with 12 GB

Cards and machines in our table with exactly 12 GB of memory. Macs use unified memory, so fewer models fit than on a 12 GB discrete GPU; open the Mac page for the default macOS limits.

RTX 3060 12GBRTX 3080 12GBRTX 4070RTX 4070 SuperRTX 4070 TiRTX 5070RX 7700 XTArc B580

Models that fit, largest first

Memory needed at 8K context. "Tight" means the model uses 90–100% of usable VRAM. How we calculate.

ModelParamsQ4_K_M4-bit fitQ8_08-bit fit
Kimi VL A3B Thinking 250616B MoE10.8 GBTight16.8 GBNo
Kimi VL A3B Instruct16B MoE10.8 GBTight16.8 GBNo
Kimi VL A3B Thinking16B MoE10.3 GBTight17.2 GBNo
LLaDA2.1 mini16B MoE10.3 GBTight17.1 GBNo
Moonlight 16B A3B Instruct16B MoE10.8 GBTight16.8 GBNo
DeepSeek Coder V2 Lite Instruct16B MoE10.5 GBTight16.4 GBNo
DeepSeek V2 Lite16B MoE9.8 GBFits16.4 GBNo
DeepSeek V2 Lite Chat16B MoE10.5 GBTight16.4 GBNo
Qwen2.5 Coder 14B Instruct15B10.6 GBTight16.8 GBNo
DeepSeek R1 Distill Qwen 14B15B10.6 GBTight16.8 GBNo
Qwen3 14B15B10.3 GBTight16.6 GBNo
phi 415B10.4 GBTight16.7 GBNo
gemma 3 12b it12B8.5 GBFits13.4 GBNo
Llama Guard 4 12B12B9.1 GBFits14.2 GBNo
Llama 3.2 11B Vision Instruct11B7.4 GBFits11.6 GBNo
Llama 3.2 11B Vision11B8.0 GBFits12.4 GBNo
GLM 4.6V Flash10B6.7 GBFits10.3 GBTight
GLM 4.1V 9B Thinking10B6.7 GBFits10.3 GBTight
Step3 VL 10B10B6.5 GBFits9.9 GBFits
SenseNova Loop 9B Preview9.7B6.4 GBFits10.0 GBFits
Qwen3.5 9B9.7B6.4 GBFits10.0 GBFits
Ornith 1.5 9B9.7B6.6 GBFits10.0 GBFits
Qwen3.5 9B Base9.7B6.4 GBFits10.0 GBFits
MiMo V2.6 Distill Qwen 9B9.4B6.6 GBFits10.0 GBFits
internlm3 8b instruct8.8B6.0 GBFits9.7 GBFits
granite 4.2 8b8.8B6.8 GBFits10.5 GBTight
granite 4.1 8b8.8B6.8 GBFits10.5 GBTight
Qwen3 VL 8B Instruct8.8B6.5 GBFits9.9 GBFits
Molmo2 8B8.7B6.5 GBFits9.9 GBFits
LFM2.5 8B A1B8.5B MoE5.5 GBFits9.1 GBFits
Qwen2.5 VL 7B Instruct8.3B5.5 GBFits8.7 GBFits
UI TARS 1.5 7B8.3B5.5 GBFits8.7 GBFits
olmOCR 2 7B 10258.3B5.5 GBFits8.7 GBFits
olmOCR 7B 0225 preview8.3B5.5 GBFits8.7 GBFits
UI TARS 7B DPO8.3B5.5 GBFits8.7 GBFits
Qwen3 8B8.2B6.5 GBFits9.9 GBFits
DeepSeek R1 0528 Qwen3 8B8.2B6.5 GBFits9.9 GBFits
granite 3.3 8b instruct8.2B6.3 GBFits9.8 GBFits
granite 3.1 8b instruct8.2B6.3 GBFits9.8 GBFits
granite 3.2 8b instruct8.2B6.3 GBFits9.8 GBFits
Apertus 8B Instruct 25098.1B6.4 GBFits9.6 GBFits
Llama 3.1 8B Instruct8.0B6.2 GBFits9.6 GBFits
Llama 3.1 8B8.0B6.2 GBFits9.6 GBFits
Hermes 3 Llama 3.1 8B8.0B6.2 GBFits9.6 GBFits
DeepSeek R1 Distill Llama 8B8.0B6.2 GBFits9.6 GBFits
Meta Llama 3 8B Instruct8.0B6.2 GBFits9.6 GBFits
Meta Llama Guard 2 8B8.0B6.2 GBFits9.6 GBFits
Meta Llama 3.1 8B Instruct8.0B6.2 GBFits9.6 GBFits
Ling 3.0 tiny7.9B MoE5.5 GBFits8.7 GBFits
MiMo 7B RL7.8B6.2 GBFits9.4 GBFits
internlm2 5 7b chat7.7B6.0 GBFits9.2 GBFits
internlm2 chat 7b7.7B6.0 GBFits9.2 GBFits
Qwen2.5 7B Instruct7.6B5.5 GBFits8.7 GBFits
Qwen2.5 Coder 7B Instruct7.6B5.5 GBFits8.7 GBFits
DeepSeek R1 Distill Qwen 7B7.6B5.5 GBFits8.7 GBFits
Falcon H1R 7B7.6B5.3 GBFits8.5 GBFits
Olmo 3 7B Instruct7.3B7.3 GBFits10.3 GBTight
Olmo 3 1025 7B7.3B7.3 GBFits10.3 GBTight
Olmo 3 7B Think7.3B7.3 GBFits10.3 GBTight
Mistral 7B v0.17.2B5.5 GBFits8.6 GBFits
Mistral 7B Instruct v0.17.2B5.5 GBFits8.6 GBFits
falcon 7b7.2B9.6 GBFits12.1 GBNo
falcon 7b instruct7.2B9.4 GBFits11.8 GBNo
granite 4.0 h tiny6.9B MoE4.6 GBFits7.5 GBFits
OLMoE 1B 7B 09246.9B MoE5.4 GBFits8.4 GBFits
deepseek coder 7b instruct v1.56.9B8.3 GBFits11.2 GBTight
deepseek coder 6.7b instruct6.7B8.3 GBFits11.1 GBTight
deepseek coder 6.7b base6.7B8.3 GBFits11.1 GBTight
Llama 2 7b hf6.7B8.3 GBFits11.1 GBTight
Yi 6B 200K6.1B4.4 GBFits7.0 GBFits
Yi 6B6.1B4.4 GBFits7.0 GBFits
Molmo2 4B4.9B4.3 GBFits6.2 GBFits
WeVisOne VL 4B4.8B4.3 GBFits6.2 GBFits
Mage VL Cowork4.7B4.3 GBFits6.2 GBFits
Qianfan OCR4.7B4.4 GBFits6.2 GBFits
Qwen3.5 4B4.7B3.7 GBFits5.3 GBFits
S5 nano4.7B3.7 GBFits5.4 GBFits
LightOnOCR 3 4B4.5B3.6 GBFits5.3 GBFits
SingVL S1 4B4.5B3.6 GBFits5.3 GBFits
Intern Decision 4B4.5B3.6 GBFits5.3 GBFits
Qwen3 VL 4B Instruct4.4B4.1 GBFits5.8 GBFits
gemma 3 4b it4.3B3.5 GBFits5.0 GBFits
medgemma 1.5 4b it4.3B3.5 GBFits5.0 GBFits
medgemma 4b it4.3B3.5 GBFits5.0 GBFits
gemma 3 4b pt4.3B3.6 GBFits5.4 GBFits
Phi 3.5 vision instruct4.1B5.8 GBFits7.6 GBFits
Qwen3 4B4.0B4.1 GBFits5.8 GBFits
Qwen3 4B Instruct 25074.0B4.1 GBFits5.8 GBFits
granite vision 4.1 4b4.0B3.2 GBFits4.6 GBFits
Phi 4 mini instruct3.8B4.1 GBFits5.6 GBFits
Phi 3 mini 4k instruct3.8B5.7 GBFits7.2 GBFits
Phi 3.5 mini instruct3.8B5.7 GBFits7.2 GBFits
Phi 3 mini 128k instruct3.8B5.7 GBFits7.2 GBFits
Qwen2.5 VL 3B Instruct3.8B2.8 GBFits4.0 GBFits
granite 4.2 3b3.7B3.3 GBFits4.8 GBFits
granite 4.1 3b llmon preview3.7B3.3 GBFits4.8 GBFits
Raon 1 nano3.5B3.0 GBFits4.3 GBFits
granite 4.1 3b3.4B3.2 GBFits4.6 GBFits
granite 4.0 micro3.4B3.2 GBFits4.6 GBFits
DeepSeek OCR 23.4B MoE2.9 GBFits4.0 GBFits
deepseek vl2 tiny3.4B MoE3.0 GBFits4.5 GBFits
Unlimited OCR 23.3B MoE3.0 GBFits4.0 GBFits
Unlimited OCR3.3B MoE2.9 GBFits4.0 GBFits
DeepSeek OCR3.3B MoE2.9 GBFits4.0 GBFits
granite 3.1 3b a800m instruct3.3B MoE2.9 GBFits4.3 GBFits
Llama 3.2 3B Instruct3.2B3.4 GBFits4.7 GBFits
Llama 3.2 3B3.2B3.6 GBFits5.1 GBFits
Hermes 3 Llama 3.2 3B3.2B3.4 GBFits4.7 GBFits
d1 3B3.1B2.3 GBFits3.4 GBFits
LFM2.5 VL 3B3.1B2.3 GBFits3.4 GBFits
Qwen2.5 3B Instruct3.1B2.9 GBFits4.3 GBFits
SmolLM3 3B3.1B3.0 GBFits4.3 GBFits
SmolLM3 3B Base3.1B3.0 GBFits4.3 GBFits
phi 22.8B4.7 GBFits5.8 GBFits
LFM2.5 2.6B2.7B2.3 GBFits3.4 GBFits
granite 3.3 2b instruct2.5B2.6 GBFits3.6 GBFits
MiniCPM5 2B2.5B2.5 GBFits3.5 GBFits
WeVisOne VL 2B2.4B2.8 GBFits3.6 GBFits
Qwen3.5 2B2.3B2.2 GBFits2.8 GBFits
SmolVLM Instruct2.2B3.0 GBFits3.8 GBFits
Qwen3 1.7B2.0B2.6 GBFits3.3 GBFits
Qwen3.8 27B DSpine b162.0B2.2 GBFits3.1 GBFits
d1 2B1.9B1.8 GBFits2.5 GBFits
LFM3 VL 2B1.9B1.8 GBFits2.5 GBFits
DeepSeek R1 Distill Qwen 1.5B1.8B1.9 GBFits2.7 GBFits
LFM3 1.7B1.7B1.8 GBFits2.5 GBFits
LFM3 1.7B Base1.7B1.7 GBFits2.5 GBFits
SmolLM2 1.7B Instruct1.7B3.0 GBFits3.7 GBFits
SmolLM2 1.7B1.7B3.0 GBFits3.7 GBFits
SmolLM 1.7B1.7B3.0 GBFits3.7 GBFits
granite 4.0 1b1.6B2.2 GBFits2.8 GBFits
Qwen2.5 1.5B Instruct1.5B1.9 GBFits2.7 GBFits
OLMo 2 0425 1B1.5B2.5 GBFits3.1 GBFits
deepseek coder 1.3b instruct1.3B2.8 GBFits3.3 GBFits
Llama 3.2 1B1.2B1.8 GBFits2.4 GBFits
Llama 3.2 1B Instruct1.2B1.6 GBFits2.1 GBFits
LFM2.5 1.2B Instruct1.2B1.3 GBFits1.8 GBFits
HunyuanOCR1.1B2.0 GBFits2.5 GBFits
PaddleOCR VL 1.6959M1.3 GBFits1.7 GBFits
Qwen3.5 0.8B873M1.5 GBFits1.7 GBFits
LightOnOCR 3 0.8B853M1.5 GBFits1.7 GBFits
FastVLM 0.5B759M1.2 GBFits1.5 GBFits
Qwen3 0.6B752M1.9 GBFits2.2 GBFits
GOT OCR2 0716M1.8 GBFits2.1 GBFits
Nemotron 3.5 Parse631M1.1 GBFits1.5 GBFits
GOT OCR 2.0 hf561M1.8 GBFits2.0 GBFits
SmolVLM2 500M Video Instruct507M1.1 GBFits1.2 GBFits
SmolVLM 500M Instruct507M1.1 GBFits1.2 GBFits
Qwen2.5 0.5B Instruct494M1.2 GBFits1.4 GBFits
LFM2.5 VL 450M449M0.8 GBFits1.0 GBFits
SmolLM2 360M Instruct362M1.1 GBFits1.2 GBFits
SmolLM2 360M362M1.1 GBFits1.2 GBFits
LFM2.5 350M354M0.8 GBFits1.0 GBFits
LFM2 350M354M1.0 GBFits1.1 GBFits

Questions

How many AI models fit in 12 GB of VRAM?

154 of the 285 open models we track fit at 4-bit (Q4_K_M) with an 8K context. 136 also fit at 8-bit (Q8_0) at the same context.

What is the largest model that fits in 12 GB?

Of the models we track, the largest that fits at 4-bit with 8K context is Kimi VL A3B Thinking 2506 (16B), needing about 10.8 GB.

How much of 12 GB can a model actually use?

About 11.4 GB. We treat 95% of a discrete GPU's VRAM as usable after driver and display reserve, and a model "fits" when it needs at most 90% of that. The formulas are on the methodology page.

Other VRAM sizes

8 GB16 GB24 GB32 GB48 GB64 GB

A generic 12 GB discrete GPU is modelled with 95% of VRAM usable, the same rule as the rest of the site. For a specific card, including its bandwidth and estimated tokens/sec, open that GPU page. Data: Hugging Face Hub API and vendor spec pages. Report an error.