Free tool · live rental prices
What AI can my PC run?
Some links on this page are referral links: we may earn a commission, at no cost to you.
Pick your graphics card, your VRAM or your Mac. You get every open model that runs on it, the ones that only run with mixture-of-experts offload, image and video models, and the cheapest GPU to rent for anything bigger.
On Nov 19, 2026, a 12 GB graphics card runs 16 of the 59 popular open models we track well at an 8K context, up to DeepSeek R1 Distill 14B and Phi-4 14B; for bigger models like Mistral Small 24B, rent an L4 from $0.03/hr (interruptible).
Check your hardware
Type your card above, or tap a memory size. Every result has its own link you can share.
Every common size, answered
What runs on 8, 12, 16, 24 GB and up
Text models at an 8K context and their own default quality (Q4_K_M, or the format they ship in), with 32 GB of system RAM for offload. Tap a model for its full memory table and rental cost, or a card for its page.
What runs on an 8 GB graphics card
Runs 9 of the open text models we track well at an 8K context, up to Qwen3 8B.
Runs great
- Llama 3.2 3B Instruct 3.5 GB
- Gemma 4 E2B 3.7 GB
- Qwen3 4B Instruct 2507 4.2 GB
- Gemma 4 E4B 5.4 GB
- Qwen2.5 7B Instruct 5.5 GB
- Mistral 7B Instruct v0.3 5.9 GB
- Llama 3.1 8B Instruct 6.3 GB
- Qwen3.5 9B 6.5 GB
- Qwen3 8B 6.5 GB
Runs with MoE offload (32 GB system RAM, slower)
- gpt-oss 20B (MoE) 14 GB
- Gemma 4 26B-A4B (MoE) 16 GB
- NVIDIA Nemotron 3 Nano 30B-A3B (MoE) 19 GB
- Qwen3 30B-A3B (MoE) 19 GB
- GLM-4.7-Flash (MoE) 19 GB
- Qwen3-VL 30B-A3B Instruct (MoE) 19 GB
- Qwen3.5 35B-A3B (MoE) 21 GB
- Qwen3.6 35B-A3B (MoE) 21 GB
- Ornith-1.5 35B-A3B (MoE) 21 GB
Images: SDXL, SD 3.5 Medium, SD 3.5 Large as a 4-bit GGUF and FLUX.2 klein 4B; FLUX.1 dev as a 4-bit GGUF only just. How we sized them
Video (NVIDIA): Wan 2.2 5B, with ComfyUI's native offloading (ComfyUI's Wan 2.2 docs). Video setup guide
Too big for 8 GB, like Gemma 4 12B? Rent NVIDIA GeForce RTX 3060 12GB from $0.03/hr (interruptible) on Vast.ai.
Vast.aiWhat runs on a 12 GB graphics card
Runs 16 of the open text models we track well at an 8K context, up to Phi-4 14B.
Runs great, new at 12 GB
- Gemma 4 12B 7.7 GB
- Gemma 3 12B 8.2 GB
- Gemma 2 9B 8.6 GB
- Qwen3 14B 10 GB
- Qwen2.5-Coder 14B 11 GB
- DeepSeek R1 Distill 14B 11 GB
- Phi-4 14B 11 GB
Plus the 9 smaller models that fit 8 GB.
Images: SDXL, SD 3.5 Medium and FLUX.2 klein 4B, plus FLUX.1 dev as a 4-bit GGUF, FLUX.2 klein 9B at fp8 and SD 3.5 Large at fp8. How we sized them
Video (NVIDIA): Wan 2.2 5B, with ComfyUI's native offloading (ComfyUI's Wan 2.2 docs). Video setup guide
Too big for 12 GB, like Mistral Small 24B? Rent NVIDIA L4 from $0.03/hr (interruptible) on Vast.ai.
Vast.ai12 GB cards: NVIDIA GeForce RTX 3060 12GB · NVIDIA GeForce RTX 3080 12GB · NVIDIA GeForce RTX 3080 Ti · NVIDIA GeForce RTX 4070 · NVIDIA GeForce RTX 4070 SUPER · NVIDIA GeForce RTX 4070 Ti
What runs on a 16 GB graphics card
Runs 17 of the open text models we track well at an 8K context, up to gpt-oss 20B (MoE).
Runs great, new at 16 GB
Plus the 16 smaller models that fit 12 GB.
Images: everything a 12 GB card runs, plus FLUX.1 dev at fp8. How we sized them
Video (NVIDIA): Wan 2.2 5B, with ComfyUI's native offloading (ComfyUI's Wan 2.2 docs). Video setup guide
Too big for 16 GB, like Mistral Small 24B? Rent NVIDIA L4 from $0.03/hr (interruptible) on Vast.ai.
Vast.ai16 GB cards: AMD Radeon RX 9060 XT 16GB · AMD Radeon RX 9070 · AMD Radeon RX 9070 XT · NVIDIA GeForce RTX 4060 Ti 16GB · NVIDIA GeForce RTX 4070 Ti SUPER · NVIDIA GeForce RTX 4080
What runs on a 24 GB graphics card
Runs 38 of the open text models we track well at an 8K context, up to Qwen2.5 Coder 32B.
Runs great, new at 24 GB
- Mistral Small 24B 15 GB
- Gemma 4 26B-A4B (MoE) 16 GB
- Devstral Small 2 24B 16 GB
- Gemma 3 27B 17 GB
- Qwen3.8 27B 17 GB
- Qwen3.5 27B 17 GB
- Qwen3.6 27B 17 GB
- NVIDIA Nemotron 3 Nano 30B-A3B (MoE) 19 GB
- Qwen3 30B-A3B (MoE) 19 GB
- Qwen3-Coder 30B-A3B (MoE) 19 GB
- GLM-4.7-Flash (MoE) 19 GB
- Gemma 2 27B 19 GB
- Gemma 4 31B 19 GB
- Qwen3-VL 30B-A3B Instruct (MoE) 19 GB
- Qwen3.5 35B-A3B (MoE) 21 GB
- Qwen3.6 35B-A3B (MoE) 21 GB
- Ornith-1.5 35B-A3B (MoE) 21 GB
- Qwen3 32B 21 GB
- DeepSeek R1 Distill 32B 21 GB
- Qwen2.5 32B Instruct 21 GB
- Qwen2.5 Coder 32B 21 GB
Plus the 17 smaller models that fit 16 GB.
Runs with MoE offload (32 GB system RAM, slower)
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Video (NVIDIA): Wan 2.2 14B at 640x640 with fp8 weights (Runpod's Wan 2.2 guide). Video setup guide
Too big for 24 GB, like Mixtral 8x7B (MoE)? Rent NVIDIA A100 40GB PCIe from $0.17/hr (interruptible) on Vast.ai.
Vast.ai24 GB cards: AMD Radeon RX 7900 XTX · NVIDIA GeForce RTX 3090 · NVIDIA GeForce RTX 3090 Ti · NVIDIA GeForce RTX 4090 · NVIDIA L4 · NVIDIA RTX A5000
What runs on a 32 GB graphics card
Runs 39 of the open text models we track well at an 8K context, up to Mixtral 8x7B (MoE).
Runs great, new at 32 GB
Plus the 38 smaller models that fit 24 GB.
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Video (NVIDIA): Wan 2.2 14B at 640x640 with fp8 weights (Runpod's Wan 2.2 guide). Video setup guide
Too big for 32 GB, like Llama 3.3 70B Instruct? Rent NVIDIA A40 from $0.17/hr (interruptible) on Vast.ai.
Vast.ai32 GB cards: NVIDIA GeForce RTX 5090 · NVIDIA RTX 5000 Ada Generation · NVIDIA RTX PRO 4500 Blackwell · NVIDIA Tesla V100 32GB PCIe · NVIDIA Tesla V100 32GB SXM2
What runs on a 48 GB graphics card
Runs 41 of the open text models we track well at an 8K context, up to DeepSeek R1 Distill 70B.
Runs great, new at 48 GB
Plus the 39 smaller models that fit 32 GB.
Runs with MoE offload (32 GB system RAM, slower)
- gpt-oss 120B (MoE) 62 GB
- Llama 4 Scout 17B-16E (MoE) 62 GB
- Mistral Small 4 119B (MoE) 68 GB
- NVIDIA Nemotron 3 Super 120B-A12B (MoE) 71 GB
- NVIDIA Nemotron 3.5 Super VL 120B-A12B (MoE) 71 GB
- Qwen3.5 122B-A10B (MoE) 72 GB
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Video (NVIDIA): Wan 2.2 14B in fp16, or in fp8 at 720p (Runpod's Wan 2.2 guide). Video setup guide
Too big for 48 GB, like Qwen3-Next 80B-A3B (MoE)? Rent NVIDIA A100 80GB SXM from $0.20/hr (interruptible) on Vast.ai.
Vast.ai48 GB cards: AMD Radeon PRO W7900 · NVIDIA A40 · NVIDIA L40 · NVIDIA L40S · NVIDIA RTX 6000 Ada Generation · NVIDIA RTX A6000
What runs on an 80 GB graphics card
Runs 50 of the open text models we track well at an 8K context, up to Qwen3.5 122B-A10B (MoE).
Runs great, new at 80 GB
- Qwen3-Coder-Next 80B-A3B (MoE) 46 GB
- Qwen3-Next 80B-A3B (MoE) 46 GB
- GLM-4.5-Air (MoE) 62 GB
- gpt-oss 120B (MoE) 62 GB
- Llama 4 Scout 17B-16E (MoE) 62 GB
- Mistral Small 4 119B (MoE) 68 GB
- NVIDIA Nemotron 3 Super 120B-A12B (MoE) 71 GB
- NVIDIA Nemotron 3.5 Super VL 120B-A12B (MoE) 71 GB
- Qwen3.5 122B-A10B (MoE) 72 GB
Plus the 41 smaller models that fit 48 GB.
Runs with MoE offload (32 GB system RAM, slower)
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Video (NVIDIA): Wan 2.2 14B at 720p without offloading (Runpod's Wan 2.2 guide). Video setup guide
Too big for 80 GB, like MiniMax-M2.5 (MoE)? Rent AMD Instinct MI300X from $2.39/hr on Runpod.
Runpod80 GB cards: NVIDIA A100 80GB PCIe · NVIDIA A100 80GB SXM · NVIDIA H100 PCIe · NVIDIA H100 SXM
What runs on a 96 GB graphics card
Runs 50 of the open text models we track well at an 8K context, up to Qwen3.5 122B-A10B (MoE).
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Video (NVIDIA): Wan 2.2 14B at 720p without offloading (Runpod's Wan 2.2 guide). Video setup guide
Too big for 96 GB, like MiniMax-M2.5 (MoE)? Rent AMD Instinct MI300X from $2.39/hr on Runpod.
Runpod96 GB cards: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition · NVIDIA RTX PRO 6000 Blackwell Server Edition · NVIDIA RTX PRO 6000 Blackwell Workstation Edition
What runs on a 128 GB graphics card
Runs 51 of the open text models we track well at an 8K context, up to Qwen3.8-Flash-Next (MoE).
Runs great, new at 128 GB
Plus the 50 smaller models that fit 96 GB.
Runs with MoE offload (32 GB system RAM, slower)
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Video (NVIDIA): Wan 2.2 14B at 720p without offloading (Runpod's Wan 2.2 guide). Video setup guide
Too big for 128 GB, like MiniMax-M2.5 (MoE)? Rent AMD Instinct MI300X from $2.39/hr on Runpod.
RunpodMacs and mini PCs
What runs on unified memory
Apple M-series Macs, DGX Spark and Strix Halo machines share one memory pool between the CPU, the GPU and the OS. We leave 8GB for macOS (or Windows/Linux) and your apps and fit models into the rest. Speed depends on memory bandwidth, which is lower than a high-end graphics card's; see Mac Studio vs DGX Spark vs Strix Halo.
What runs on a 32 GB unified memory (Mac, DGX Spark, Strix Halo)
Runs 38 of the open text models we track well at an 8K context, up to Qwen2.5 Coder 32B. We leave 8 GB for the OS and apps, so models get about 24 GB.
Runs great
- Llama 3.2 3B Instruct 3.5 GB
- Gemma 4 E2B 3.7 GB
- Qwen3 4B Instruct 2507 4.2 GB
- Gemma 4 E4B 5.4 GB
- Qwen2.5 7B Instruct 5.5 GB
- Mistral 7B Instruct v0.3 5.9 GB
- Llama 3.1 8B Instruct 6.3 GB
- Qwen3.5 9B 6.5 GB
- Qwen3 8B 6.5 GB
- Gemma 4 12B 7.7 GB
- Gemma 3 12B 8.2 GB
- Gemma 2 9B 8.6 GB
- Qwen3 14B 10 GB
- Qwen2.5-Coder 14B 11 GB
- DeepSeek R1 Distill 14B 11 GB
- Phi-4 14B 11 GB
- gpt-oss 20B (MoE) 14 GB
- Mistral Small 24B 15 GB
- Gemma 4 26B-A4B (MoE) 16 GB
- Devstral Small 2 24B 16 GB
- Gemma 3 27B 17 GB
- Qwen3.8 27B 17 GB
- Qwen3.5 27B 17 GB
- Qwen3.6 27B 17 GB
- NVIDIA Nemotron 3 Nano 30B-A3B (MoE) 19 GB
- Qwen3 30B-A3B (MoE) 19 GB
- Qwen3-Coder 30B-A3B (MoE) 19 GB
- GLM-4.7-Flash (MoE) 19 GB
- Gemma 2 27B 19 GB
- Gemma 4 31B 19 GB
- Qwen3-VL 30B-A3B Instruct (MoE) 19 GB
- Qwen3.5 35B-A3B (MoE) 21 GB
- Qwen3.6 35B-A3B (MoE) 21 GB
- Ornith-1.5 35B-A3B (MoE) 21 GB
- Qwen3 32B 21 GB
- DeepSeek R1 Distill 32B 21 GB
- Qwen2.5 32B Instruct 21 GB
- Qwen2.5 Coder 32B 21 GB
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Too big for 32 GB, like Mixtral 8x7B (MoE)? Rent NVIDIA A100 40GB PCIe from $0.17/hr (interruptible) on Vast.ai.
Vast.aiWhat runs on a 64 GB unified memory (Mac, DGX Spark, Strix Halo)
Runs 43 of the open text models we track well at an 8K context, up to Qwen3-Next 80B-A3B (MoE). We leave 8 GB for the OS and apps, so models get about 56 GB.
Runs great, new at 64 GB
- Mixtral 8x7B (MoE) 28 GB
- Llama 3.3 70B Instruct 43 GB
- DeepSeek R1 Distill 70B 43 GB
- Qwen3-Coder-Next 80B-A3B (MoE) 46 GB
- Qwen3-Next 80B-A3B (MoE) 46 GB
Plus the 38 smaller models that fit 32 GB.
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Too big for 64 GB, like GLM-4.5-Air (MoE)? Rent NVIDIA A100 80GB SXM from $0.20/hr (interruptible) on Vast.ai.
Vast.aiWhat runs on a 128 GB unified memory (Mac, DGX Spark, Strix Halo)
Runs 51 of the open text models we track well at an 8K context, up to Qwen3.8-Flash-Next (MoE). We leave 8 GB for the OS and apps, so models get about 120 GB.
Runs great, new at 128 GB
- GLM-4.5-Air (MoE) 62 GB
- gpt-oss 120B (MoE) 62 GB
- Llama 4 Scout 17B-16E (MoE) 62 GB
- Mistral Small 4 119B (MoE) 68 GB
- NVIDIA Nemotron 3 Super 120B-A12B (MoE) 71 GB
- NVIDIA Nemotron 3.5 Super VL 120B-A12B (MoE) 71 GB
- Qwen3.5 122B-A10B (MoE) 72 GB
- Qwen3.8-Flash-Next (MoE) 103 GB
Plus the 43 smaller models that fit 64 GB.
Images: everything a 16 GB card runs, plus FLUX.2 dev at 4-bit, and FLUX.2 klein 9B and SD 3.5 Large at full precision. How we sized them
Too big for 128 GB, like MiniMax-M2.5 (MoE)? Rent AMD Instinct MI300X from $2.39/hr on Runpod.
RunpodHow we calculate it
- Memory a model needs = weights at its default quality + the KV cache for the context + about 0.8 GB of runtime overhead, from each model's own configuration on Hugging Face. Same engine as our VRAM calculator.
- Runs great = it needs at most 90% of the memory, so your desktop and a browser still fit.
- Runs with MoE offload = a mixture-of-experts model whose active weights fit on the card while its expert weights sit in system RAM (llama.cpp's --n-cpu-moe), with 8 GB of RAM kept for the OS. It works, but slower.
- Images come from the model files on Hugging Face (our rule); video from ComfyUI's and Runpod's Wan 2.2 figures (NVIDIA cards).
- Rental = the cheapest live offer with enough memory across the providers we track, captured Thu, 19 Nov 2026 04:20:07 UTC. Refreshed every hour. Spot offers can be interrupted.
- These are estimates from model files and memory math, not benchmarks. Full method: how we calculate.