Best GPU for running long-memory models locally

Which GPU or Mac you need for open models built for long-term memory (MemT, Hy-Memory), from the cheapest card that runs the small one to what it takes for MemT-1.0-Flash. All checks are at Q4_K_M, about 4k context.

Cheapest way in: Intel Arc Pro B70 32GB runs MemT-1.0-Nano for $949. For MemT-1.0-Flash, the cheapest fit we track is Mac Studio M5 Ultra 512GB at $22299.

Cheapest hardware per model

"Fits" means the model needs 90% or less of usable memory at Q4 (how we calculate). Prices are dated snapshots, mostly launch MSRP, so street prices can differ a lot; see each device page for its source.

Every priced machine, cheapest first

We may earn a commission when you buy through these links, at no extra cost to you.

Other guides

• Best GPU for running Llama locally3 models
• Best GPU for running Qwen3 locally4 models
• Best GPU for running gpt-oss locally2 models
• Best GPU for running DeepSeek R1 locally4 models
• Best GPU for running Gemma 3 locally3 models
• Best GPU for running Gemma 4 locally5 models
• Best GPU for running Qwen3.8 locally3 models
• Best GPU for running Ornith 1.5 locally3 models
• Best GPU for running Nemotron locally3 models
• Best GPU for running Liquid AI LFM locally2 models
• Best GPU for running Mistral locally4 models
• Best GPU for running GLM locally4 models
• Best GPU for running Phi locally4 models
• Best GPU for running local coding models locally8 models