• Specifiche del produttore, niente a occhio
  • Ogni macchina abbinata ai modelli che può eseguire
  • Spedizione in tutto il mondo, dazi e tasse inclusi

How much memory each local model size needs, and which machines and cards hold it

Last checked 21 November 2026

This reference page is available in English only.

Short answer

A local language model needs about 6 GB of memory for a 7–8B model at 4-bit, about 20 GB for a 30–34B model and about 42 GB for a 70B model; at 8-bit the same 70B model needs about 80 GB. A 70B model at 4-bit fits on a 48 GB graphics card or a unified-memory machine with 64 GB or more, and 120B-class models need a 96 GB card or 128 GB of unified memory.

The table below is built from the same figures and rules as our model guide, so the two always agree. It covers the 18 graphics cards, accelerators and unified-memory systems we list today. Prices change, so they are on each product page, not here.

Memory needed by model size

Approximate figures, last checked 21 November 2026. Rules of thumb, not measurements; see how they are worked out below.
Model sizeQuantizationMemory neededSmallest that holds itOur products that hold it
7–8BLlama 3.1 8B, Mistral 7B, Qwen2.5 7B4-bitabout 6 GB7 GB card8 GB unified12 GB: GIGABYTE GeForce RTX 5070 WINDFORCE OC SFF 12G16 GB: GIGABYTE GeForce RTX 5070 Ti WINDFORCE OC SFF 16G, GIGABYTE GeForce RTX 5080 WINDFORCE OC SFF 16G, Intel Arc Pro B5024 GB: ASRock Intel Arc Pro B60 Creator 24GB, NVIDIA GeForce RTX 4090, SAPPHIRE PULSE AMD Radeon RX 7900 XTX 24GB, Tenstorrent Wormhole n300s24 GB unified: Apple Mac mini (M4 Pro)32 GB: NVIDIA GeForce RTX 5090, PowerColor AMD Radeon AI PRO R9700 32GB, Tenstorrent Blackhole p150a48 GB: AMD Radeon PRO W7900 (48 GB)96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)96 GB unified: Apple Mac Studio (M3 Ultra)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
7–8BLlama 3.1 8B, Mistral 7B, Qwen2.5 7B8-bitabout 10 GB12 GB card14 GB unified12 GB: GIGABYTE GeForce RTX 5070 WINDFORCE OC SFF 12G16 GB: GIGABYTE GeForce RTX 5070 Ti WINDFORCE OC SFF 16G, GIGABYTE GeForce RTX 5080 WINDFORCE OC SFF 16G, Intel Arc Pro B5024 GB: ASRock Intel Arc Pro B60 Creator 24GB, NVIDIA GeForce RTX 4090, SAPPHIRE PULSE AMD Radeon RX 7900 XTX 24GB, Tenstorrent Wormhole n300s24 GB unified: Apple Mac mini (M4 Pro)32 GB: NVIDIA GeForce RTX 5090, PowerColor AMD Radeon AI PRO R9700 32GB, Tenstorrent Blackhole p150a48 GB: AMD Radeon PRO W7900 (48 GB)96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)96 GB unified: Apple Mac Studio (M3 Ultra)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
13–14BPhi-4 (14B), Qwen2.5 14B4-bitabout 10 GB12 GB card14 GB unified12 GB: GIGABYTE GeForce RTX 5070 WINDFORCE OC SFF 12G16 GB: GIGABYTE GeForce RTX 5070 Ti WINDFORCE OC SFF 16G, GIGABYTE GeForce RTX 5080 WINDFORCE OC SFF 16G, Intel Arc Pro B5024 GB: ASRock Intel Arc Pro B60 Creator 24GB, NVIDIA GeForce RTX 4090, SAPPHIRE PULSE AMD Radeon RX 7900 XTX 24GB, Tenstorrent Wormhole n300s24 GB unified: Apple Mac mini (M4 Pro)32 GB: NVIDIA GeForce RTX 5090, PowerColor AMD Radeon AI PRO R9700 32GB, Tenstorrent Blackhole p150a48 GB: AMD Radeon PRO W7900 (48 GB)96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)96 GB unified: Apple Mac Studio (M3 Ultra)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
13–14BPhi-4 (14B), Qwen2.5 14B8-bitabout 16 GB18 GB card22 GB unified24 GB: ASRock Intel Arc Pro B60 Creator 24GB, NVIDIA GeForce RTX 4090, SAPPHIRE PULSE AMD Radeon RX 7900 XTX 24GB, Tenstorrent Wormhole n300s24 GB unified: Apple Mac mini (M4 Pro)32 GB: NVIDIA GeForce RTX 5090, PowerColor AMD Radeon AI PRO R9700 32GB, Tenstorrent Blackhole p150a48 GB: AMD Radeon PRO W7900 (48 GB)96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)96 GB unified: Apple Mac Studio (M3 Ultra)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
30–34BGemma 2 27B, Qwen2.5 32B4-bitabout 20 GB23 GB card27 GB unified24 GB: ASRock Intel Arc Pro B60 Creator 24GB, NVIDIA GeForce RTX 4090, SAPPHIRE PULSE AMD Radeon RX 7900 XTX 24GB, Tenstorrent Wormhole n300s32 GB: NVIDIA GeForce RTX 5090, PowerColor AMD Radeon AI PRO R9700 32GB, Tenstorrent Blackhole p150a48 GB: AMD Radeon PRO W7900 (48 GB)96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)96 GB unified: Apple Mac Studio (M3 Ultra)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
30–34BGemma 2 27B, Qwen2.5 32B8-bitabout 38 GB43 GB card51 GB unified48 GB: AMD Radeon PRO W7900 (48 GB)96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)96 GB unified: Apple Mac Studio (M3 Ultra)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
70BLlama 3.3 70B, Qwen2.5 72B4-bitabout 42 GB47 GB card56 GB unified48 GB: AMD Radeon PRO W7900 (48 GB)96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)96 GB unified: Apple Mac Studio (M3 Ultra)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
70BLlama 3.3 70B, Qwen2.5 72B8-bitabout 80 GB89 GB card107 GB unified96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
120B+gpt-oss-120b (MoE), Mistral Large 2 (123B)4-bitabout 80 GB89 GB card107 GB unified96 GB: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB)128 GB unified: Beelink GTR9 Pro, Minisforum MS-S1 Max, NVIDIA DGX Spark
120B+gpt-oss-120b (MoE), Mistral Large 2 (123B)8-bitabout 140 GB156 GB card187 GB unifiedNone of our single products. Two or more cards together can hold it; ask for a custom build.

For one product at a time, the model guide shows the same result with live prices for your market.

How these numbers are worked out

  1. Weights. A model's weights take about parameters × bits ÷ 8 bytes. A 70B model at 4 bits is about 35 GB of weights.
  2. Headroom. We add 20–30% for the context (the KV cache), runtime buffers and the operating system's share, and round to a conservative figure for the whole size class. The 70B class covers 65B to 72B models and is counted at about 42 GB at 4-bit.
  3. Usable memory. On a graphics card or accelerator we treat 90% of its memory as usable. On a unified-memory system (Apple silicon, Ryzen AI Max+, DGX Spark) we treat 75% as usable, because the operating system keeps part of the shared pool and GPU limits apply.
  4. Fit. A product holds a model size when its usable memory is at least the figure in the table. The minimum column is that figure divided by 90% or 75%, rounded up to a whole gigabyte.
  5. Products. The product column lists the graphics cards, accelerators and unified-memory systems we list today that pass that check, smallest memory first. It is the same list the model guide shows, and it changes when a product goes out of stock or comes back.

Limits of this table

  • These figures are rules of thumb worked out from model sizes. They are not measurements of each model on each machine.
  • Long contexts need more memory, because the KV cache grows with every token the model keeps in view. Serving several users at once needs more again.
  • Image and video models use memory differently from text models; plan more headroom for them.
  • Mixture-of-experts (MoE) models run faster than dense models of the same size, but all of their weights still have to be in memory, so they need the full figure.
  • The table covers what fits, not how fast it runs. Once a model fits, writing speed depends mostly on memory bandwidth.
  • Each graphics card is counted on its own. Two or more cards in one machine add their memory together, but splitting a model across them does not double the speed.
  • Products with several memory options are counted at their base configuration: the Mac mini at 24 GB and the Mac Studio at 96 GB. Larger configurations hold more.
  • On Ryzen AI Max+ machines under Windows, the GPU can be given up to 96 GB of the 128 GB; Linux lets you assign more.

Product specifications are the makers' own figures, from our catalog. Last checked 21 November 2026.

Questions people ask

Can a 128 GB mini PC run a 70B model?

Yes. A 70B model needs about 42 GB at 4-bit and about 80 GB at 8-bit, and we count 75% of 128 GB of unified memory, about 96 GB, as usable, so both fit. On Windows the GPU of a Ryzen AI Max+ 395 machine can be given up to 96 GB; Linux lets you assign more. See Ryzen AI Max+ 395 mini PCs compared.

How much VRAM do I need for a 30B model at 4-bit?

About 20 GB for the 30–34B size class. With 90% of a card's memory counted as usable, that means a card with at least 23 GB, so a 24 GB or 32 GB card. At 8-bit the same model needs about 38 GB, more than any single 32 GB card holds.

Is 32 GB on an RTX 5090 enough for a 70B model?

No. A 70B model needs about 42 GB at 4-bit, and about 28.8 GB of the card's 32 GB is usable. One RTX 5090 runs models up to 30–34B at 4-bit. For 70B, plan for two cards, the 48 GB Radeon PRO W7900, the 96 GB RTX PRO 6000 or a unified-memory machine with 64 GB or more.

DGX Spark, Mac Studio or Ryzen AI Max+ 395 for local LLMs?

The DGX Spark and the Ryzen AI Max+ 395 mini PCs have 128 GB of unified memory and hold a 70B model at 8-bit or a 120B-class model at 4-bit. The Mac Studio with M3 Ultra comes with 96 GB as standard, enough for a 70B model at 4-bit, and moves 819 GB/s of memory bandwidth against the Spark's 273 GB/s, so large dense models generate text faster on it. Choose the Spark for CUDA development, the Mac Studio for speed in a quiet box on macOS, and a Ryzen AI Max+ machine for 128 GB on Windows or Linux.

What does unified memory mean for model size?

A unified-memory machine shares one pool of memory between the CPU and the GPU, so the GPU can use far more than a graphics card has. The operating system keeps part of the pool, so we count 75% as usable, against 90% on a graphics card. A Mac mini with 24 GB has about 18 GB for models: enough for a 14B model at 8-bit, not for a 30B model at 4-bit.

What can a 16 GB graphics card run?

Models up to 13–14B at 4-bit (about 10 GB) and up to 7–8B at 8-bit (about 10 GB). About 14.4 GB of the 16 GB is usable, which is not enough for a 30B model at 4-bit, at about 20 GB. The RTX 5070 Ti, the RTX 5080 and the Intel Arc Pro B50 all fall in this group.

Do mixture-of-experts models need less memory?

No. A mixture-of-experts (MoE) model uses only part of its weights for each token, so it runs faster than a dense model of the same size, but all of its weights still have to be in memory. Plan for the full figure of its size class: gpt-oss-120b, for example, is in the 120B+ class.

Can I add two graphics cards together for a bigger model?

Yes. Splitting a model across two cards adds their memory together, but it does not double the speed, and it needs two suitable slots, more power and more airflow. Two 32 GB cards, for example, give 64 GB, enough for a 70B model at 4-bit at about 42 GB. One larger card is simpler when the budget allows.

Read more