Llama Models: Model Sizes and RAM for Local AI

Meta's Llama is the most widely used open-weight model family in the world. The Llama 4 generation is multimodal MoE: Scout at 109B total / 17B active with 10M context, and Maverick at 400B total / 17B active, both of which want a 96GB+ Mac Studio rather than a laptop. The 3.x sizes remain the everyday local picks, from 1B and 3B on small machines up to 8B, 70B and 405B as memory allows. With the largest ecosystem of fine-tunes, tools, and community support, Llama is the safest default choice for local AI.

developer
Meta
local models
12
size range
1B-405B
RAM range
2-320 GB
Table 1Local builds

All Llama models

12 local models
ModelSizeQuantVRAMMin RAMBest ForQualityOllama
Llama 3.2 1B Instruct1BQ4_K_M1 GB2 GBChat48
Llama 3.2 3B Instruct3BQ4_K_M2.5 GB4 GBChat66
Llama 3.1 8B Instruct8BQ4_K_M6.5 GB12 GBChat, Coding77
Llama 3.1 8B Instruct (Q8)8BQ8_08 GB16 GBChat, Coding79
Llama 3.1 8B Instruct (Q5)8BQ5_K_M8 GB12 GBChat, Coding79
Llama 3.1 70B Instruct70BQ4_K_M42 GB64 GBQuality, Coding84
Llama 3.3 70B Instruct70BQ4_K_M42 GB64 GBQuality, Coding85
Llama 3.3 70B Instruct (Q8)70BQ8_075 GB96 GBQuality, Coding87
Llama 3.3 70B Instruct (Q6)70BQ6_K57.9 GB96 GBQuality, Coding86
Llama 4 Scout109BQ4_K_M67 GB96 GBLong context, Quality, Multimodal85
Llama 4 Maverick400BQ4_K_M245 GB320 GBFrontier quality, Long context87
Llama 3.1 405B Instruct405BQ4_K_M243 GB320 GBQuality, Reasoning, Coding87
4-bit builds

Llama model sizes in GB

8 builds

Every local Llama model at 4-bit, from 1B to 405B parameters. Load is our estimate of the memory the model takes once loaded (est.). The file size is the same quant as listed on Hugging Face, where we track it.

ModelParamsQuantQ4 load (est.)File on HFMin RAM
Llama 3.2 1B Instruct1BQ4_K_M1 GB-2 GB
Llama 3.2 3B Instruct3BQ4_K_M2.5 GB-4 GB
Llama 3.1 8B Instruct8BQ4_K_M6.5 GB4.58 GB12 GB
Llama 3.1 70B Instruct70BQ4_K_M42 GB-64 GB
Llama 3.3 70B Instruct70BQ4_K_M42 GB-64 GB
Llama 4 Scout109BQ4_K_M67 GB60.87 GB96 GB
Llama 4 Maverick400BQ4_K_M245 GB-320 GB
Llama 3.1 405B Instruct405BQ4_K_M243 GB-320 GB

Llama 3.2 1B RAM requirements

Llama 3.2 1B loads in about 1 GB at Q4_K_M (est.) and needs at least 2 GB of RAM.

Llama 3.2 3B RAM requirements

Llama 3.2 3B loads in about 2.5 GB at Q4_K_M (est.) and needs at least 4 GB of RAM.

Llama 3.1 8B RAM requirements

Llama 3.1 8B loads in about 6.5 GB at Q4_K_M (est.) and needs at least 12 GB of RAM.

By memory

Best local Llama model by RAM

For each amount of RAM, the Llama build with the highest ModelFit quality score whose minimum RAM fits. On a tie, the larger model wins.

RAMBest Llama modelQuantLoad (est.)Min RAMQuality
8 GBLlama 3.2 3B InstructQ4_K_M2.5 GB4 GB66
16 GBLlama 3.1 8B Instruct (Q8)Q8_08 GB16 GB79
24 GBLlama 3.1 8B Instruct (Q8)Q8_08 GB16 GB79
32 GBLlama 3.1 8B Instruct (Q8)Q8_08 GB16 GB79
64 GBLlama 3.3 70B InstructQ4_K_M42 GB64 GB85
128 GBLlama 3.3 70B Instruct (Q8)Q8_075 GB96 GB87

Key features

  • Most popular open-weight model family
  • Strong general reasoning and instruction following
  • Llama 4 Scout: 109B MoE with 10M context
  • Llama 3.x sizes from 1B to 405B for any machine
Table 2

Device compatibility

Which Llama models can run on each device class, based on minimum RAM requirements.

ModeliPhoneAirProStudioMini
Llama 3.2 1B Instruct (1B)ExcellentExcellentExcellentExcellentExcellentExcellent
Llama 3.2 3B Instruct (3B)ExcellentExcellentExcellentExcellentExcellentExcellent
Llama 3.1 8B Instruct (8B)PossiblePossiblePossibleExcellentPossiblePossible
Llama 3.1 8B Instruct (Q8) (8B)NoPossiblePossibleExcellentPossiblePossible
Llama 3.1 8B Instruct (Q5) (8B)PossiblePossiblePossibleExcellentPossiblePossible
Llama 3.1 70B Instruct (70B)NoNoPossiblePossiblePossibleNo
Llama 3.3 70B Instruct (70B)NoNoPossiblePossiblePossibleNo
Llama 3.3 70B Instruct (Q8) (70B)NoNoPossiblePossibleNoNo
Llama 3.3 70B Instruct (Q6) (70B)NoNoPossiblePossibleNoNo
Llama 4 Scout (109B)NoNoPossiblePossibleNoNo
Llama 4 Maverick (400B)NoNoNoPossibleNoNo
Llama 3.1 405B Instruct (405B)NoNoNoPossibleNoNo

RAM requirements

Fig. 1Estimated load per model, with the minimum RAM
Llama 3.2 1B Instruct1 GB · min 2 GB
Llama 3.2 3B Instruct2.5 GB · min 4 GB
Llama 3.1 8B Instruct6.5 GB · min 12 GB
Llama 3.1 8B Instruct (Q8)8 GB · min 16 GB
Llama 3.1 8B Instruct (Q5)8 GB · min 12 GB
Llama 3.1 70B Instruct42 GB · min 64 GB
Llama 3.3 70B Instruct42 GB · min 64 GB
Llama 3.3 70B Instruct (Q8)75 GB · min 96 GB
Llama 3.3 70B Instruct (Q6)57.9 GB · min 96 GB
Llama 4 Scout67 GB · min 96 GB
Llama 4 Maverick245 GB · min 320 GB
Llama 3.1 405B Instruct243 GB · min 320 GB

Config reports

No speed test run for this family yet. Run the speed test: npx @modelfit/cli@latest bench

Share your config

Every report is reviewed before it appears. No email or account needed.

0/280

Frequently asked questions

What is the best Llama model for a MacBook?

Llama 3.2 3B for MacBook Air (8GB RAM) or Llama 3.1 8B for MacBook Pro (16GB+ RAM). The 8B model is the community favorite for general-purpose local AI.

Can Llama 70B run on a Mac?

Yes, but you need at least 48GB RAM (Mac Studio or maxed-out MacBook Pro). The Q4 quantized version uses about 42GB. Expect an estimated ~10 tokens per second on an M4 Max with 128GB. The newer Llama 4 Scout is far heavier: 67GB loaded, so a 96GB or 128GB Mac Studio.

What is the difference between Llama 3.1 and 3.2?

Llama 3.2 added small sizes (1B, 3B) optimized for edge devices and mobile. Llama 3.1 covers 8B, 70B, and 405B. For most users, pick 3.2 3B for small devices or 3.1 8B for laptops.

How does Llama compare to Qwen?

Llama has stronger general reasoning and a larger community. Qwen has more size options and better multilingual support. At 7-8B, they are close in quality. Pick based on your use case.

Getting started

  1. 01Best LLM for MacBook →
  2. 02How to Set Up Ollama →
  3. 03Browse All Devices →
  4. 04ModelFit Wizard →