How much VRAM do I need for a 30B–32B model?

About 23.8 GB at Q4_K_M, 41.0 GB at Q8_0 and 76.4 GB at FP16, at a ~4k context. That is for Qwen3-32B (32.8B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.

How much VRAM does a 30B–32B model need for 32k context?

About 31.4 GB at Q4_K_M with an FP16 KV cache, or 27.1 GB with a Q8 cache, based on Qwen3-32B's published config. Models with different attention layouts need more or less.

What is the cheapest GPU for a 30B–32B model?

Of the GPUs we track, the Intel Arc Pro B70 32GB (32 GB, $949 list price) is the cheapest that runs it fully on the GPU at Q4.

Can a Mac run a 30B–32B model?

Yes. The cheapest Mac or unified-memory mini PC we track that runs it at Q4 is the Mac mini M5 Pro 64GB (64 GB, about 48.0 GB usable, $3,199).

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 30B–32B model?

Short answer: about 23.8 GB at Q4_K_M, the usual quant for local use. Below: every quant, longer contexts, the cheapest hardware that fits, and every 30B–32B open model we track.

Q4_K_M: 23.8 GB 19.9 GB weights
Q8_0: 41.0 GB 34.8 GB weights
FP16: 76.4 GB 65.5 GB weights

Worked out for Qwen3-32B (32.8B 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: Intel Arc Pro B70 32GB (32 GB, $949). Cheapest Mac or mini PC: Mac mini M5 Pro 64GB ($3,199).

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 30B–32B model at each quant
Setting	Memory needed	Cheapest GPU	Cheapest Mac / mini PC
Q4_K_M, ~4k context	23.8 GB	Intel Arc Pro B70 32GB 32 GB usable · $949	Mac mini M5 Pro 64GB 48 GB usable · $3,199
Q8_0, ~4k context	41.0 GB	NVIDIA RTX A6000 48 GB usable · $4,649	Mac mini M5 Pro 64GB 48 GB usable · $3,199
FP16, ~4k context	76.4 GB	NVIDIA RTX Pro 6000 Blackwell 96GB 96 GB usable · $8,565	Framework Desktop (Ryzen AI Max+ 395) 128GB 96 GB usable · $3,449
Q4_K_M, 32k context	31.4 GB	NVIDIA RTX A6000 48 GB usable · $4,649	Mac mini M5 Pro 64GB 48 GB usable · $3,199

Memory by context length

Each extra token of context adds to the KV cache. Numbers for Qwen3-32B, from its published config; a Q8 KV cache halves the cache part.

Memory needed for a 30B–32B model by context length
Context	Q4_K_M, FP16 cache	Q4_K_M, Q8 cache	Q8_0, FP16 cache
4k tokens	23.8 GB	23.8 GB	41.0 GB
8k tokens	24.9 GB	23.8 GB	42.1 GB
16k tokens	27.1 GB	24.9 GB	44.3 GB
32k tokens	31.4 GB	27.1 GB	48.5 GB
64k tokens	39.9 GB	31.4 GB	57.1 GB
128k tokens	57.1 GB	39.9 GB	74.3 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 30B–32B model at Q4
Hardware	Q4 need vs. usable	Q4 / Q8 / FP16	Est. tok/s
NVIDIA Tesla P40 24GBGPU	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	10–15
Radeon RX 7900 XTXGPU · $999	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	29–41
GeForce RTX 3090GPU · $1,499	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	28–40
GeForce RTX 4090GPU · $1,599	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	30–43
Mac mini M6 32GBunified memory · $1,799	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	4–7
GeForce RTX 3090 TiGPU · $1,999	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	30–43
NVIDIA RTX A5000GPU · $2,250	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	23–33
MacBook Pro M6 32GBunified memory · $2,399	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	4–7
NVIDIA Titan RTXGPU · $2,499	23.8 / 24.0 GB	Q4TightQ8Won’t fitFP16Won’t fit	20–29
NVIDIA Tesla V100 32GB (PCIe)GPU	23.8 / 32.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	27–39
AMD Radeon Instinct MI50 32GBGPU	23.8 / 32.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	31–44
AMD Radeon Instinct MI60 32GBGPU	23.8 / 32.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	31–44
Intel Arc Pro B70 32GBGPU · $949	23.8 / 32.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	18–26
GeForce RTX 5090GPU · $1,999	23.8 / 32.0 GB	Q4FitsQ8Won’t fitFP16Won’t fit	54–77
Mac mini M5 Pro 64GBunified memory · $3,199	23.8 / 48.0 GB	Q4FitsQ8FitsFP16Won’t fit	7–12
NVIDIA RTX A6000GPU · $4,649	23.8 / 48.0 GB	Q4FitsQ8FitsFP16Won’t fit	23–33
NVIDIA DGX Spark 64GBunified memory · $4,999	23.8 / 48.0 GB	Q4FitsQ8FitsFP16Won’t fit	6–11
NVIDIA RTX 6000 Ada 48GBGPU · $6,799	23.8 / 48.0 GB	Q4FitsQ8FitsFP16Won’t fit	29–41
Framework Desktop (Ryzen AI Max+ 395) 128GBunified memory · $3,449	23.8 / 96.0 GB	Q4FitsQ8FitsFP16Fits	6–10
MacBook Pro M4 Max 128GBunified memory · $4,999	23.8 / 96.0 GB	Q4FitsQ8FitsFP16Fits	12–22
MacBook Pro M5 Max 128GBunified memory · $6,699	23.8 / 96.0 GB	Q4FitsQ8FitsFP16Fits	14–25
NVIDIA DGX Spark 128GBunified memory · $6,950	23.8 / 96.0 GB	Q4FitsQ8FitsFP16Fits	6–11
NVIDIA RTX Pro 6000 Blackwell 96GBGPU · $8,565	23.8 / 96.0 GB	Q4FitsQ8FitsFP16Fits	54–77
Mac Studio M5 Ultra 256GBunified memory · $9,499	23.8 / 192.0 GB	Q4FitsQ8FitsFP16Fits	27–48
Mac Studio M5 Ultra 512GBunified memory · $22,299	23.8 / 384.0 GB	Q4FitsQ8FitsFP16Fits	27–48

30B–32B models we track

Memory needed by each 30B–32B model we track
Model	Q4_K_M	Q8_0	FP16
ERNIE-4.5-VL-28B-A3B-Thinking29.7B params · baidu · MoE	21.7 GB	37.2 GB	69.2 GB
Muse-Glimmer-30B29.8B params · meta-models	21.8 GB	37.4 GB	69.5 GB
granite-switch-4.1-30b30.0B params · ibm-granite	21.9 GB	37.6 GB	70.0 GB
Hy-Memory-30B-A3B30.1B params · tencent · MoE	22.0 GB	37.7 GB	70.1 GB
Hy-MT2-30B-A3B30.1B params · tencent · MoE	22.0 GB	37.7 GB	70.1 GB
North-Mini-Code-1.030.5B params · CohereLabs · MoE	22.3 GB	38.2 GB	71.1 GB
Qwen3-30B-A3B30.5B params · Qwen · MoE	22.3 GB	38.3 GB	71.2 GB
Qwen3-Coder-30B-A3B-Instruct30.5B params · Qwen · MoE	22.3 GB	38.3 GB	71.2 GB
Qwen3-30B-A3B-Instruct-250730.5B params · Qwen · MoE	22.3 GB	38.3 GB	71.2 GB
Qwen3-30B-A3B-Thinking-250730.5B params · Qwen · MoE	22.3 GB	38.3 GB	71.2 GB
Tongyi-DeepResearch-30B-A3B30.5B params · Alibaba-NLP · MoE	22.3 GB	38.3 GB	71.2 GB
Rubicon-Preview30.5B params · inclusionAI · MoE	22.3 GB	38.3 GB	71.2 GB
kanana-2-30b-a3b-instruct-260130.7B params · kakaocorp · MoE	22.4 GB	38.5 GB	71.5 GB
kanana-2-30b-a3b-instruct30.7B params · kakaocorp · MoE	22.4 GB	38.5 GB	71.5 GB
kanana-2-30b-a3b-thinking-260130.7B params · kakaocorp · MoE	22.4 GB	38.5 GB	71.5 GB
Qwen3-VL-30B-A3B-Instruct31.1B params · Qwen · MoE	22.7 GB	39.0 GB	72.5 GB
Qwen3-VL-30B-A3B-Thinking31.1B params · Qwen · MoE	22.7 GB	39.0 GB	72.5 GB
UI-Venus-1.5-30B-A3B31.1B params · inclusionAI · MoE	22.7 GB	39.0 GB	72.5 GB
Xing4.0-29B-A4B31.2B params · XingChen-AGI · MoE	22.8 GB	39.1 GB	72.8 GB
Xing4.0-29B-Intent31.2B params · XingChen-AGI · MoE	22.8 GB	39.1 GB	72.8 GB
GLM-4.7-Flash31.2B params · zai-org · MoE	22.8 GB	39.1 GB	72.8 GB
Gemma 4 31B31.3B params · google	22.8 GB	39.2 GB	72.9 GB
gemma-4-31B31.3B params · google	22.8 GB	39.2 GB	72.9 GB
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF1631.6B params · nvidia · MoE	23.0 GB	39.6 GB	73.6 GB
NVIDIA-Nemotron-3-Nano-30B-A3B-BF1631.6B params · nvidia · MoE	23.0 GB	39.6 GB	73.6 GB
Nemotron-Cascade-2-30B-A3B31.6B params · nvidia · MoE	23.0 GB	39.6 GB	73.6 GB
EXAONE-3.5-32B-Instruct32.0B params · LGAI-EXAONE	23.3 GB	40.1 GB	74.6 GB
EXAONE-Deep-32B32.0B params · LGAI-EXAONE	23.3 GB	40.1 GB	74.6 GB
EXAONE-4.0-32B32.0B params · LGAI-EXAONE	23.3 GB	40.1 GB	74.6 GB
EXAONE-4.0.1-32B32.0B params · LGAI-EXAONE	23.3 GB	40.1 GB	74.6 GB
sarvam-30b32.2B params · sarvamai · MoE	23.4 GB	40.3 GB	75.0 GB
granite-4.0-h-small32.2B params · ibm-granite · MoE	23.5 GB	40.4 GB	75.1 GB
Olmo-3-32B-Think32.2B params · allenai	23.5 GB	40.4 GB	75.1 GB
Olmo-3-1125-32B32.2B params · allenai	23.5 GB	40.4 GB	75.1 GB
Olmo-3.1-32B-Instruct32.2B params · allenai	23.5 GB	40.4 GB	75.1 GB
Olmo-3.1-32B-Think32.2B params · allenai	23.5 GB	40.4 GB	75.1 GB
Olmo-3.1-32B-Instruct-DPO32.2B params · allenai	23.5 GB	40.4 GB	75.1 GB
OLMo-2-0325-32B-Instruct32.2B params · allenai	23.5 GB	40.4 GB	75.1 GB
OLMo-2-0325-32B32.2B params · allenai	23.5 GB	40.4 GB	75.1 GB
GLM-Z1-32B-041432.6B params · zai-org	23.7 GB	40.8 GB	75.9 GB
GLM-4-32B-041432.6B params · zai-org	23.7 GB	40.8 GB	75.9 GB
Qwen3-32B32.8B params · Qwen	23.8 GB	41.0 GB	76.4 GB
Qwen2.5-Coder-32B-Instruct32.8B params · Qwen	23.8 GB	41.0 GB	76.4 GB
DeepSeek-R1-Distill-Qwen-32B32.8B params · deepseek-ai	23.8 GB	41.0 GB	76.4 GB
Qwen2.5-32B-Instruct32.8B params · Qwen	23.8 GB	41.0 GB	76.4 GB
QwQ-32B32.8B params · Qwen	23.8 GB	41.0 GB	76.4 GB
QwQ-32B-Preview32.8B params · Qwen	23.8 GB	41.0 GB	76.4 GB
Qwen2.5-32B32.8B params · Qwen	23.8 GB	41.0 GB	76.4 GB
Qwen2.5-Coder-32B32.8B params · Qwen	23.8 GB	41.0 GB	76.4 GB
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF1633.0B params · nvidia · MoE	24.0 GB	41.3 GB	76.9 GB
Nanbeige4.3-33B33.1B params · Nanbeige	24.1 GB	41.5 GB	77.2 GB
GLM-Z1-Rumination-32B-041433.1B params · zai-org	24.1 GB	41.5 GB	77.2 GB
FlexOlmo-7x7B-1T33.3B params · allenai	24.2 GB	41.7 GB	77.5 GB
HyperCLOVAX-SEED-Think-32B33.3B params · naver-hyperclovax	24.2 GB	41.7 GB	77.6 GB
Qwen3-VL-32B-Instruct33.4B params · Qwen	24.3 GB	41.8 GB	77.7 GB
Laguna-XS.233.4B params · poolside · MoE	24.3 GB	41.9 GB	77.9 GB
Qwen2.5-VL-32B-Instruct33.5B params · Qwen	24.3 GB	41.9 GB	77.9 GB
Falcon-H1-34B-Instruct33.6B params · tiiuae	24.5 GB	42.1 GB	78.4 GB
EXAONE-4.5-33B34.4B params · LGAI-EXAONE	24.9 GB	43.0 GB	80.0 GB
Qwen-AgentWorld-35B-A3B34.7B params · Qwen · MoE	25.2 GB	43.4 GB	80.7 GB
KAT-Coder-V2.5-Dev34.7B params · Kwaipilot · MoE	25.2 GB	43.4 GB	80.7 GB
Clover-35B-A3B-Preview34.7B params · Diffuspace · MoE	25.2 GB	43.4 GB	80.7 GB
Agents-A135.1B params · InternScience · MoE	25.5 GB	43.9 GB	81.7 GB
Nex-N2.5-mini35.1B params · nex-agi · MoE	25.5 GB	43.9 GB	81.7 GB
Nex-N2-mini35.1B params · nex-agi · MoE	25.5 GB	43.9 GB	81.7 GB
Axion-1.0-35B35.1B params · Axion1-0 · MoE	25.5 GB	43.9 GB	81.7 GB
NorthCUA-Mini35.1B params · beipo-ai · MoE	25.5 GB	43.9 GB	81.7 GB
MemT-1.0-Nano35.1B params · MemTensor · MoE	25.5 GB	43.9 GB	81.8 GB
MiniCPM-V-4.7-35B-A3B35.2B params · openbmb · MoE	25.5 GB	44.0 GB	82.0 GB
Ornith-1.5-35B-A3B36.0B params · ornith-ai · MoE	26.1 GB	44.9 GB	83.7 GB
Sage-A1.536.0B params · InternScience	26.1 GB	44.9 GB	83.7 GB
Qwen3.6-35B-A3B36.0B params · Qwen · MoE	26.1 GB	44.9 GB	83.7 GB
Qwen3.5-35B-A3B36.0B params · Qwen · MoE	26.1 GB	44.9 GB	83.7 GB
Apex-35B-V1.036.0B params · Apexintelligence-AI · MoE	26.1 GB	44.9 GB	83.7 GB
Intern-S2-Preview36.1B params · internlm · MoE	26.2 GB	45.1 GB	84.0 GB
Seed-OSS-36B-Instruct36.2B params · ByteDance-Seed	26.2 GB	45.2 GB	84.1 GB
Seed-OSS-36B-Base-woSyn36.2B params · ByteDance-Seed	26.2 GB	45.2 GB	84.1 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 30B–32B model?: About 23.8 GB at Q4_K_M, 41.0 GB at Q8_0 and 76.4 GB at FP16, at a ~4k context. That is for Qwen3-32B (32.8B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.
How much VRAM does a 30B–32B model need for 32k context?: About 31.4 GB at Q4_K_M with an FP16 KV cache, or 27.1 GB with a Q8 cache, based on Qwen3-32B's published config. Models with different attention layouts need more or less.
What is the cheapest GPU for a 30B–32B model?: Of the GPUs we track, the Intel Arc Pro B70 32GB (32 GB, $949 list price) is the cheapest that runs it fully on the GPU at Q4.
Can a Mac run a 30B–32B model?: Yes. The cheapest Mac or unified-memory mini PC we track that runs it at Q4 is the Mac mini M5 Pro 64GB (64 GB, about 48.0 GB usable, $3,199).
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 · 12B–14B · 16B–21B · 24B–27B · 40B–50B · 70B · 120B · 200B–300B · 300B–480B · 500B–800B · 1T+ · All sizes · Ollama VRAM by tag · GGUF sizes by quant