How much VRAM do I need for a 12B–14B model?

About 11.3 GB at Q4_K_M, 19.0 GB at Q8_0 and 35.0 GB at FP16, at a ~4k context. That is for Qwen3-14B (14.8B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.

How much VRAM does a 12B–14B model need for 32k context?

About 16.0 GB at Q4_K_M with an FP16 KV cache, or 13.3 GB with a Q8 cache, based on Qwen3-14B's published config. Models with different attention layouts need more or less.

What is the cheapest GPU for a 12B–14B model?

Of the GPUs we track, the Radeon RX 7600 XT (16 GB, $329 list price) is the cheapest that runs it fully on the GPU at Q4.

Can a Mac run a 12B–14B model?

Yes. The cheapest Mac or unified-memory mini PC we track that runs it at Q4 is the Mac mini M6 24GB (24 GB, about 18.0 GB usable, $1,299).

Does a MoE model of the same size need less VRAM?

No. A mixture-of-experts model still has to hold all of its weights in memory, so it needs about the same VRAM as a dense model with the same total parameters. It runs faster, because only a few experts are read per token.

How much VRAM do you need for a 12B–14B model?

Short answer: about 11.3 GB at Q4_K_M, the usual quant for local use. Below: every quant, longer contexts, the cheapest hardware that fits, and every 12B–14B open model we track.

Q4_K_M: 11.3 GB 9.0 GB weights
Q8_0: 19.0 GB 15.7 GB weights
FP16: 35.0 GB 29.5 GB weights

Worked out for Qwen3-14B (14.8B parameters, from the Hugging Face API). Formula: weights × 1.15 + 1 GB at a ~4k context (how we calculate).

Cheapest GPU that fits at Q4: Radeon RX 7600 XT (16 GB, $329). Cheapest Mac or mini PC: Mac mini M6 24GB ($1,299).

Cheapest hardware for each quant

Only devices with a sourced price. GPUs at list price (MSRP); street prices may be higher.

Cheapest GPU and Mac or mini PC for a 12B–14B model at each quant
Setting	Memory needed	Cheapest GPU	Cheapest Mac / mini PC
Q4_K_M, ~4k context	11.3 GB	Radeon RX 7600 XT 16 GB usable · $329	Mac mini M6 24GB 18 GB usable · $1,299
Q8_0, ~4k context	19.0 GB	Intel Arc Pro B70 32GB 32 GB usable · $949	Mac mini M6 32GB 24 GB usable · $1,799
FP16, ~4k context	35.0 GB	NVIDIA RTX A6000 48 GB usable · $4,649	Mac mini M5 Pro 64GB 48 GB usable · $3,199
Q4_K_M, 32k context	16.0 GB	Radeon RX 7900 XT 20 GB usable · $899	Mac mini M6 24GB 18 GB usable · $1,299

Memory by context length

Each extra token of context adds to the KV cache. Numbers for Qwen3-14B, from its published config; a Q8 KV cache halves the cache part.

Memory needed for a 12B–14B model by context length
Context	Q4_K_M, FP16 cache	Q4_K_M, Q8 cache	Q8_0, FP16 cache
4k tokens	11.3 GB	11.3 GB	19.0 GB
8k tokens	12.0 GB	11.3 GB	19.7 GB
16k tokens	13.3 GB	12.0 GB	21.1 GB
32k tokens	16.0 GB	13.3 GB	23.7 GB
64k tokens	21.4 GB	16.0 GB	29.1 GB
128k tokens	32.1 GB	21.4 GB	39.8 GB

Hardware that runs it

Bar: Q4 memory needed vs. usable memory. Speed is estimated decode at Q4 (how we estimate).FitsTight

Tracked hardware that runs a 12B–14B model at Q4
Hardware	Q4 need vs. usable	Q4 / Q8 / FP16	Est. tok/s
GeForce RTX 2060 12GBGPU	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	23–32
Intel Arc B580GPU · $249	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	31–43
GeForce RTX 3060 12GBGPU · $329	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	24–34
Radeon RX 7700 XTGPU · $449	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	29–41
Radeon RX 6700 XTGPU · $479	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	26–36
GeForce RTX 5070GPU · $549	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	45–64
Radeon RX 6750 XTGPU · $549	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	29–41
GeForce RTX 4070GPU · $599	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	34–48
GeForce RTX 4070 SuperGPU · $599	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	34–48
Mac mini M4 16GBunified memory · $789	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	6–11
GeForce RTX 3080 12GBGPU · $799	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	61–87
GeForce RTX 4070 TiGPU · $799	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	34–48
Mac mini M6 16GBunified memory · $899	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	8–14
GeForce RTX 3080 TiGPU · $1,199	11.3 / 12.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	61–87
Radeon RX 7600 XTGPU · $329	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	19–27
Radeon RX 9060 XT 16GBGPU · $349	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	21–30
Intel Arc A770 16GBGPU · $349	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	38–53
GeForce RTX 5060 Ti 16GBGPU · $429	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	30–43
GeForce RTX 4060 Ti 16GBGPU · $499	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	19–27
Radeon RX 7800 XTGPU · $499	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	42–59
Radeon RX 7900 GREGPU · $549	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	39–55
Radeon RX 9070GPU · $549	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	43–61
Radeon RX 6800GPU · $579	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	34–49
Radeon RX 9070 XTGPU · $599	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	43–61
Radeon RX 6800 XTGPU · $649	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	34–49
NVIDIA RTX 2000 Ada 16GBGPU · $649	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	15–21
GeForce RTX 5070 TiGPU · $749	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	60–85
GeForce RTX 4070 Ti SuperGPU · $799	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	45–64
GeForce RTX 4080 SuperGPU · $999	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	49–70
GeForce RTX 5080GPU · $999	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	64–91
Radeon RX 6900 XTGPU · $999	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	34–49
NVIDIA RTX A4000GPU · $1,000	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	30–43
Radeon RX 6950 XTGPU · $1,099	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	39–55
GeForce RTX 4080GPU · $1,199	11.3 / 16.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	48–68
Mac mini M6 24GBunified memory · $1,299	11.3 / 18.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	9–15
Radeon RX 7900 XTGPU · $899	11.3 / 20.0 GB	Q4FitsQ8TightFP16Won’t fit	54–76
NVIDIA RTX 4000 Ada 20GBGPU · $1,250	11.3 / 20.0 GB	Q4FitsQ8TightFP16Won’t fit	24–34
NVIDIA Tesla P40 24GBGPU	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	23–33
Radeon RX 7900 XTXGPU · $999	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	64–91
GeForce RTX 3090GPU · $1,499	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	63–89
GeForce RTX 4090GPU · $1,599	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	68–96
Mac mini M6 32GBunified memory · $1,799	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	9–15
GeForce RTX 3090 TiGPU · $1,999	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	68–96
NVIDIA RTX A5000GPU · $2,250	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	51–73
MacBook Pro M6 32GBunified memory · $2,399	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	9–15
NVIDIA Titan RTXGPU · $2,499	11.3 / 24.0 GB	Q4FitsQ8FitsFP16Won’t fit	45–64
NVIDIA Tesla V100 32GB (PCIe)GPU	11.3 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	60–85
AMD Radeon Instinct MI50 32GBGPU	11.3 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	69–97
AMD Radeon Instinct MI60 32GBGPU	11.3 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	69–97
Intel Arc Pro B70 32GBGPU · $949	11.3 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	41–58
GeForce RTX 5090GPU · $1,999	11.3 / 32.0 GB	Q4FitsQ8FitsFP16Won’t fit	120–170
Mac mini M5 Pro 64GBunified memory · $3,199	11.3 / 48.0 GB	Q4FitsQ8FitsFP16Fits	15–27
NVIDIA RTX A6000GPU · $4,649	11.3 / 48.0 GB	Q4FitsQ8FitsFP16Fits	51–73
NVIDIA DGX Spark 64GBunified memory · $4,999	11.3 / 48.0 GB	Q4FitsQ8FitsFP16Fits	14–24
NVIDIA RTX 6000 Ada 48GBGPU · $6,799	11.3 / 48.0 GB	Q4FitsQ8FitsFP16Fits	64–91
Framework Desktop (Ryzen AI Max+ 395) 128GBunified memory · $3,449	11.3 / 96.0 GB	Q4FitsQ8FitsFP16Fits	13–23
MacBook Pro M4 Max 128GBunified memory · $4,999	11.3 / 96.0 GB	Q4FitsQ8FitsFP16Fits	27–49
MacBook Pro M5 Max 128GBunified memory · $6,699	11.3 / 96.0 GB	Q4FitsQ8FitsFP16Fits	31–55
NVIDIA DGX Spark 128GBunified memory · $6,950	11.3 / 96.0 GB	Q4FitsQ8FitsFP16Fits	14–24
NVIDIA RTX Pro 6000 Blackwell 96GBGPU · $8,565	11.3 / 96.0 GB	Q4FitsQ8FitsFP16Fits	120–170
Mac Studio M5 Ultra 256GBunified memory · $9,499	11.3 / 192.0 GB	Q4FitsQ8FitsFP16Fits	60–107
Mac Studio M5 Ultra 512GBunified memory · $22,299	11.3 / 384.0 GB	Q4FitsQ8FitsFP16Fits	60–107

12B–14B models we track

Memory needed by each 12B–14B model we track
Model	Q4_K_M	Q8_0	FP16
Gemma 4 12B12.0B params · google	9.3 GB	15.6 GB	28.5 GB
Mellum2.1-12B-A2.5B-Thinking12.1B params · JetBrains · MoE	9.5 GB	15.8 GB	28.9 GB
Mellum2-12B-A2.5B-Thinking12.1B params · JetBrains · MoE	9.5 GB	15.8 GB	28.9 GB
Mellum2-12B-A2.5B-Instruct12.1B params · JetBrains · MoE	9.5 GB	15.8 GB	28.9 GB
Mellum2-12B-A2.5B-Thinking-SFT12.1B params · JetBrains · MoE	9.5 GB	15.8 GB	28.9 GB
Mellum2.1-12B-A2.5B-Thinking-MTP12.1B params · JetBrains · MoE	9.5 GB	15.8 GB	28.9 GB
gemma-3-12b-it12.2B params · google	9.5 GB	15.9 GB	29.0 GB
gemma-3-12b-pt12.2B params · google	9.5 GB	15.9 GB	29.0 GB
Mistral-Nemo-Instruct-240712.2B params · mistralai	9.5 GB	16.0 GB	29.2 GB
OLMo-2-1124-13B-Instruct13.7B params · allenai	10.6 GB	17.8 GB	32.5 GB
phi-414.7B params · microsoft	11.2 GB	18.9 GB	34.7 GB
Phi-4-reasoning14.7B params · microsoft	11.2 GB	18.9 GB	34.7 GB
Phi-4-reasoning-plus14.7B params · microsoft	11.2 GB	18.9 GB	34.7 GB
HyperCLOVAX-SEED-Think-14B14.7B params · naver-hyperclovax	11.3 GB	19.0 GB	34.9 GB
Qwen3-14B14.8B params · Qwen	11.3 GB	19.0 GB	35.0 GB
JanusCoder-14B14.8B params · internlm	11.3 GB	19.0 GB	35.0 GB
Qwen2.5-Coder-14B-Instruct14.8B params · Qwen	11.3 GB	19.0 GB	35.0 GB
DeepSeek-R1-Distill-Qwen-14B14.8B params · deepseek-ai	11.3 GB	19.0 GB	35.0 GB
Qwen2.5-14B-Instruct14.8B params · Qwen	11.3 GB	19.0 GB	35.0 GB
Qwen2.5-14B14.8B params · Qwen	11.3 GB	19.0 GB	35.0 GB
Qwen2.5-14B-Instruct-1M14.8B params · Qwen	11.3 GB	19.0 GB	35.0 GB
Qwen2.5-Coder-14B14.8B params · Qwen	11.3 GB	19.0 GB	35.0 GB
Carbon-X (DNA)14.8B params · HuggingFaceBio	11.3 GB	19.1 GB	35.1 GB
Apriel-1.5-15b-Thinker14.9B params · ServiceNow-AI	11.4 GB	19.2 GB	35.2 GB
Apriel-1.6-15b-Thinker14.9B params · ServiceNow-AI	11.4 GB	19.2 GB	35.2 GB
Apriel-Nemotron-15b-Thinker15.0B params · ServiceNow-AI	11.4 GB	19.3 GB	35.4 GB
Phi-4-reasoning-vision-15B15.1B params · microsoft	11.5 GB	19.5 GB	35.8 GB
kanana-1.5-15.7b-a3b-instruct15.7B params · kakaocorp · MoE	11.9 GB	20.2 GB	37.1 GB
DeepSeek-Coder-V2-Lite-Instruct15.7B params · deepseek-ai · MoE	12.0 GB	20.2 GB	37.1 GB
Instella-MoE-16B-A3B-Think15.9B params · amd · MoE	12.1 GB	20.4 GB	37.5 GB
Moonlight-16B-A3B-Instruct16.0B params · moonshotai · MoE	12.1 GB	20.5 GB	37.7 GB
Moonlight-16B-A3B16.0B params · moonshotai · MoE	12.1 GB	20.5 GB	37.7 GB

Have a different amount of memory? Try the VRAM calculator or see best LLMs by VRAM size.

Quick answers

How much VRAM do I need for a 12B–14B model?: About 11.3 GB at Q4_K_M, 19.0 GB at Q8_0 and 35.0 GB at FP16, at a ~4k context. That is for Qwen3-14B (14.8B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.
How much VRAM does a 12B–14B model need for 32k context?: About 16.0 GB at Q4_K_M with an FP16 KV cache, or 13.3 GB with a Q8 cache, based on Qwen3-14B's published config. Models with different attention layouts need more or less.
What is the cheapest GPU for a 12B–14B model?: Of the GPUs we track, the Radeon RX 7600 XT (16 GB, $329 list price) is the cheapest that runs it fully on the GPU at Q4.
Can a Mac run a 12B–14B model?: Yes. The cheapest Mac or unified-memory mini PC we track that runs it at Q4 is the Mac mini M6 24GB (24 GB, about 18.0 GB usable, $1,299).
Does a MoE model of the same size need less VRAM?: No. A mixture-of-experts model still has to hold all of its weights in memory, so it needs about the same VRAM as a dense model with the same total parameters. It runs faster, because only a few experts are read per token.

Other sizes: 1B–2B · 2B–3B · 4B–6B · 7B–8B · 16B–21B · 24B–27B · 30B–32B · 40B–50B · 70B · 120B · 200B–300B · 300B–480B · 500B–800B · 1T+ · All sizes · Ollama VRAM by tag · GGUF sizes by quant