Run it locally

Will your machine run a local AI model?

Local and open-weight models run on your own hardware. No per-token bill, but they need enough memory. Tell us what you have (or what you're thinking of buying) and see what fits.

What are you running it on?

Graphics

Memory (RAM), GB

System type

Usable for a model: ~13 GB · 16 GB system RAM (CPU inference)

Runs comfortably

· fits in memory with headroom

Ministral 3 14B · Efficient local daily-driver

needs ~10 GB · 14B

Phi-4 · Compact reasoning model

needs ~10 GB · 14B

Local model (12B via Ollama) · Local / private budget model

needs ~9 GB · 12B

GLM-4.6V · Low-latency vision-language model

needs ~8 GB · 9B

Ministral 3 8B · Small efficient local model

needs ~7 GB · 8B

Granite 4.1 8B · Enterprise/open lightweight assistant

needs ~7 GB · 8B

Qwen3-VL 8B · Small local VLM

needs ~7 GB · 8B

Won't fit

· needs more memory than you have

GPT-OSS 20B · Small open-weight reasoning/tool model

needs ~15 GB · 21B

Codestral · Coding model

needs ~15 GB · 22B

Mistral Small 3.2 · Cheap small generalist

needs ~16 GB · 24B

Gemma 4 26B A4B · Efficient multimodal MoE

needs ~18 GB · 26B

Qwen3.6 27B · Strong local generalist/coder

needs ~18 GB · 27B

Gemma 3 27B · Small open model

needs ~18 GB · 27B

Qwen3-Coder 30B · Practical local coding model

needs ~20 GB · 30B

Nemotron 3 Nano 30B · Low-latency MoE reasoning/chat

needs ~20 GB · 30B

Qwen3-VL 30B · Stronger local VLM

needs ~20 GB · 30B

Gemma 4 31B · Open multimodal local generalist

needs ~21 GB · 31B

Qwen3.6 35B A3B · Local multimodal coding/general model

needs ~23 GB · 35B

DeepSeek R1 Distill Llama 70B · Larger local reasoning model

needs ~44 GB · 70B

Llama 3.3 70B · Open-weight generalist

needs ~44 GB · 70B

Qwen3-Coder-Next · Coding-agent specialist

needs ~50 GB · 80B

GPT-OSS 120B · Open-weight GPT-style model

needs ~74 GB · 120B

Nemotron 3 Super 120B · Open / enterprise-tuned generalist

needs ~74 GB · 120B

Devstral 2 · Agentic coding model

needs ~76 GB · 123B

MiniMax M2 · Agentic coding / workflows model

needs ~140 GB · 230B

Llama 4 Maverick · Open-weight / flexible deployment model

needs ~242 GB · 400B

Devices to consider

All-in-one options, at their top memory: what each can run comfortably.

Device	Max memory	Runs up to	From
Raspberry Pi 5 ↗ CPU-only inference — small models, and slow	16 GB	Ministral 3 14B (~10 GB)	$80
Apple Mac mini (M4) ↗ Unified memory; good value entry for local AI	32 GB	Gemma 4 31B (~21 GB)	$599
Apple Mac mini (M4 Pro) ↗ Up to 64 GB unified — mid-size models	64 GB	Qwen3.6 35B A3B (~23 GB)	$1,399
Apple Mac Studio (M4 Max) ↗ Up to 128 GB unified — large models	128 GB	Devstral 2 (~76 GB)	$1,999
NVIDIA DGX Spark ↗ Grace Blackwell, purpose-built for local AI; 128 GB unified	128 GB	Devstral 2 (~76 GB)	$4,699

Build a local setup Compare local vs cloud

Memory needs are modeled at Q4 (4-bit) quantization (roughly total parameters × 0.6 GB, plus overhead): a “will it fit” estimate, not a speed benchmark. Device specs are from vendor pages (updated 2026-07-06). Running a model that barely fits will be slow.