Meta · 3.2B · Q4_K_M · 1.88 GiB

Llama 3.2 3B Instruct at Q4_K_M needs about 3.42 GiB at 8K context, and the smallest listed machine that fits it is GeForce RTX 3060 Ti.

3B dense instruct model. Architecture from the public GGUF header.

ollama run llama3.2:3b

Weights: Llama-3.2-3B-Instruct-Q4_K_M.gguf. Architecture: config source. Hugging Face repo created 2024-09-18. License field: llama3.2.

Quants with a measured file size

QuantFileSize
Q4_K_M · recommendedLlama-3.2-3B-Instruct-Q4_K_M.gguf1.88 GiB
Q5_K_MLlama-3.2-3B-Instruct-Q5_K_M.gguf2.16 GiB
Q6_KLlama-3.2-3B-Instruct-Q6_K.gguf2.46 GiB
Q8_0Llama-3.2-3B-Instruct-Q8_0.gguf3.19 GiB

Machines

Same quant and 8,192 tokens on every row.
MachineMemoryTotalVerdictSpeed
Apple M1, 8 GB8 GB unified3.42 GiBFits13–27 tok/s
Apple M2, 8 GB8 GB unified3.42 GiBFits19–40 tok/s
Apple M3, 8 GB8 GB unified3.42 GiBFits19–40 tok/s
GeForce RTX 3060 Ti8 GB GDDR3.42 GiBFits83–177 tok/s
GeForce RTX 30708 GB GDDR3.42 GiBFits83–177 tok/s
GeForce RTX 40608 GB GDDR3.42 GiBFits51–108 tok/s
GeForce RTX 4060 Ti 8GB8 GB GDDR3.42 GiBFits54–114 tok/s
GeForce RTX 50608 GB GDDR3.42 GiBFits83–177 tok/s
GeForce RTX 5060 Ti 8GB8 GB GDDR3.42 GiBFits83–177 tok/s
Radeon RX 66008 GB GDDR63.42 GiBFits42–89 tok/s
Radeon RX 76008 GB GDDR63.42 GiBFits54–114 tok/s
GeForce RTX 3080 10GB10 GB GDDR3.42 GiBFits141–301 tok/s
GeForce RTX 3060 12GB12 GB GDDR3.42 GiBFits67–143 tok/s
GeForce RTX 407012 GB GDDR3.42 GiBFits94–200 tok/s
GeForce RTX 4070 Super12 GB GDDR3.42 GiBFits94–200 tok/s
GeForce RTX 4070 Ti12 GB GDDR3.42 GiBFits94–200 tok/s
GeForce RTX 507012 GB GDDR3.42 GiBFits125–266 tok/s
Intel Arc B58012 GB GDDR63.42 GiBFits85–181 tok/s
Radeon RX 6700 XT12 GB GDDR63.42 GiBFits71–152 tok/s
Radeon RX 7700 XT12 GB GDDR63.42 GiBFits80–171 tok/s
Apple M1 Pro, 16 GB16 GB unified3.42 GiBFits37–79 tok/s
Apple M1, 16 GB16 GB unified3.42 GiBFits13–27 tok/s
Apple M2 Pro, 16 GB16 GB unified3.42 GiBFits37–79 tok/s
Apple M2, 16 GB16 GB unified3.42 GiBFits19–40 tok/s
Apple M3, 16 GB16 GB unified3.42 GiBFits19–40 tok/s
Apple M4, 16 GB16 GB unified3.42 GiBFits22–48 tok/s
GeForce RTX 4060 Ti 16GB16 GB GDDR3.42 GiBFits54–114 tok/s
GeForce RTX 4070 Ti Super16 GB GDDR3.42 GiBFits125–266 tok/s
GeForce RTX 408016 GB GDDR3.42 GiBFits133–284 tok/s
GeForce RTX 4080 Super16 GB GDDR3.42 GiBFits137–292 tok/s
GeForce RTX 5060 Ti 16GB16 GB GDDR3.42 GiBFits83–177 tok/s
GeForce RTX 5070 Ti16 GB GDDR3.42 GiBFits167–355 tok/s
GeForce RTX 508016 GB GDDR3.42 GiBFits178–380 tok/s
Intel Arc A770 16GB16 GB GDDR63.42 GiBFits95–203 tok/s
Legion Go16 GB LPDDR5X-75003.42 GiBFits22–48 tok/s
Radeon RX 6800 XT16 GB GDDR63.42 GiBFits95–203 tok/s
Radeon RX 7800 XT16 GB GDDR63.42 GiBFits116–247 tok/s
Radeon RX 9060 XT 16GB16 GB GDDR63.42 GiBFits59–127 tok/s
Radeon RX 9070 XT16 GB GDDR63.42 GiBFits119–254 tok/s
ROG Ally16 GB LPDDR5-64003.42 GiBFits19–41 tok/s
Steam Deck LCD16 GB LPDDR5-55003.42 GiBFits16–35 tok/s
Steam Deck OLED16 GB LPDDR5-64003.42 GiBFits19–41 tok/s
Apple M3 Pro, 18 GB18 GB unified3.42 GiBFits28–59 tok/s
Radeon RX 7900 XT20 GB GDDR63.42 GiBFits149–317 tok/s
Apple M2, 24 GB24 GB unified3.42 GiBFits19–40 tok/s
Apple M3, 24 GB24 GB unified3.42 GiBFits19–40 tok/s
Apple M4 Pro, 24 GB24 GB unified3.42 GiBFits51–108 tok/s
Apple M4, 24 GB24 GB unified3.42 GiBFits22–48 tok/s
GeForce RTX 309024 GB GDDR3.42 GiBFits174–371 tok/s
GeForce RTX 409024 GB GDDR3.42 GiBFits187–399 tok/s
Radeon RX 7900 XTX24 GB GDDR63.42 GiBFits178–380 tok/s
ROG Ally X24 GB LPDDR5X-75003.42 GiBFits22–48 tok/s
Apple M1 Max, 32 GB32 GB unified3.42 GiBFits74–158 tok/s
Apple M1 Pro, 32 GB32 GB unified3.42 GiBFits37–79 tok/s
Apple M2 Max, 32 GB32 GB unified3.42 GiBFits74–158 tok/s
Apple M2 Pro, 32 GB32 GB unified3.42 GiBFits37–79 tok/s
Apple M4, 32 GB32 GB unified3.42 GiBFits22–48 tok/s
GeForce RTX 509032 GB GDDR3.42 GiBFits333–710 tok/s
Apple M3 Pro, 36 GB36 GB unified3.42 GiBFits28–59 tok/s
Apple M4 Max 14-core, 36 GB36 GB unified3.42 GiBFits76–162 tok/s
Apple M3 Max 16-core, 48 GB48 GB unified3.42 GiBFits74–158 tok/s
Apple M4 Max 16-core, 48 GB48 GB unified3.42 GiBFits101–216 tok/s
Apple M4 Pro, 48 GB48 GB unified3.42 GiBFits51–108 tok/s
Apple M1 Max, 64 GB64 GB unified3.42 GiBFits74–158 tok/s
Apple M2 Max, 64 GB64 GB unified3.42 GiBFits74–158 tok/s
Apple M3 Max 16-core, 64 GB64 GB unified3.42 GiBFits74–158 tok/s
Apple M4 Max 16-core, 64 GB64 GB unified3.42 GiBFits101–216 tok/s
Apple M4 Pro, 64 GB64 GB unified3.42 GiBFits51–108 tok/s
Apple M2 Max, 96 GB96 GB unified3.42 GiBFits74–158 tok/s
Apple M3 Max 16-core, 128 GB128 GB unified3.42 GiBFits74–158 tok/s
Apple M4 Max 16-core, 128 GB128 GB unified3.42 GiBFits101–216 tok/s

Exact answers

Steam Deck OLEDfitsApple M1, 8 GBfits

FAQ

Why Q4_K_M?

Q4_K_M is the quant this page recommends. The file is Llama-3.2-3B-Instruct-Q4_K_M.gguf at 1.88 GiB. Other single-file quants are in the table when a real size was available.

Where do the architecture numbers come from?

GGUF metadata on the Q4_K_M file. Meta's config.json is access-gated. Source: https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/blob/main/Llama-3.2-3B-Instruct-Q4_K_M.gguf

Does this include a long context?

The fit uses 8,192 tokens, or less when that is all the fast memory can hold. The model file allows up to 131,072 tokens, and KV cache grows as you use more of that.

What is the Ollama command?

ollama run llama3.2:3b. The library tag is Q4-class or the vendor quant. Compare it with the GGUF linked on this page before you treat the sizes as identical.