Pick two devices. See the differences on one page.
Can I run AI locally on this device — or that one? The answer comes down to two numbers: VRAM and memory bandwidth. Pick two devices below and Cairn runs the same 50 open-weight models against each, showing which fits, at what quantization, and how fast.
Device A
Device B
Kimi K2
1000B · Moonshot AI
DeepSeek V3
685B · DeepSeek
DeepSeek V3.2
685B · DeepSeek
DeepSeek R1 671B
671B · DeepSeek
DeepSeek V3.1
671B · DeepSeek
Qwen 3 Coder 480B (MoE)
480B · Alibaba
Llama 3.1 405B
405B · Meta
Llama 4 Maverick 17B-128E
400B · Meta
Qwen 3 235B (MoE)
235B · Alibaba
Mixtral 8x22B
141B · Mistral AI
Devstral 2 123B
123B · Mistral AI
GPT-OSS 120B
117B · OpenAI
Llama 4 Scout 17B
109B · Meta
Qwen 2.5 72B
72B · Alibaba
Llama 3.1 70B
70B · Meta
Llama 3.3 70B
70B · Meta
DeepSeek R1 Distill 70B
70B · DeepSeek
Mixtral 8x7B
47B · Mistral AI
Qwen 3 32B
32B · Alibaba
Qwen 2.5 Coder 32B
32B · Alibaba
DeepSeek R1 Distill 32B
32B · DeepSeek
Qwen 3 30B-A3B (MoE)
30B · Alibaba
Gemma 3 27B
27B · Google
Mistral Small 24B
24B · Mistral AI
Mistral Small 3.1 24B
24B · Mistral AI
GPT-OSS 20B
21B · OpenAI
Phi-4 14B
14B · Microsoft
Qwen 3 14B
14B · Alibaba
DeepSeek R1 Distill 14B
14B · DeepSeek
Gemma 3 12B
12B · Google
Mistral Nemo 12B
12B · Mistral AI
Llama 3.2 11B Vision
11B · Meta
Qwen 3.5 9B
9B · Alibaba
Gemma 2 9B
9B · Google
Llama 3.1 8B
8B · Meta
Qwen 3 8B
8B · Alibaba
DeepSeek R1 Distill 7B
7B · DeepSeek
Code Llama 7B
7B · Meta
LLaVA 1.6 7B
7B · LLaVA Team
Qwen 2.5 Coder 7B
7B · Alibaba
Qwen 3 4B
4B · Alibaba
Gemma 3 4B
4B · Google
Phi-3.5 Mini 3.8B
3.8B · Microsoft
Llama 3.2 3B
3B · Meta
Qwen 3 1.7B
1.7B · Alibaba
DeepSeek R1 Distill 1.5B
1.5B · DeepSeek
Llama 3.2 1B
1B · Meta
Gemma 3 1B
1B · Google
Qwen 3.5 0.8B
0.8B · Alibaba
Qwen 3 0.6B
0.6B · Alibaba
6 GB of VRAM is the floor — enough for 7B models at Q4 quantization. 12 GB covers most 13B models; 24 GB opens up 30B; 70B needs 48 GB+ or a unified-memory setup like Mac Studio. NVIDIA RTX (discrete VRAM), Apple Silicon (unified memory), and AMD Radeon all work.
Depends on what you want. Apple wins on capacity: Mac Studio unified memory goes up to 192 GB. NVIDIA wins on bandwidth: an RTX 4090 is ~1 TB/s vs the M2 Pro's 200 GB/s. Bigger models fit on Apple; NVIDIA runs them faster.
Both, but they answer different questions. VRAM decides if a model loads. Bandwidth decides how fast it runs. An RTX 4090 has 24 GB VRAM and ~1 TB/s bandwidth — a 7B model fits in 5 GB and runs at 80+ tokens/sec. An Apple M2 Pro has 16 GB unified memory at 200 GB/s — same model fits but runs around 20 tok/sec.
Quantization. At Q4_K_M a 70B model needs ~42 GB; at Q8_0 it needs ~75 GB. Two 48 GB cards will both fit Q4, but context length and activation memory can push Q8 past the edge on one of them.