How much VRAM do I need for a 500B–800B model?

About 478.2 GB at Q4_K_M, 837.4 GB at Q8_0 and 1575.3 GB at FP16, at a ~4k context. That is for DeepSeek-R1 (684.5B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.

What is the cheapest GPU for a 500B–800B model?

None of the single consumer GPUs we track hold it at Q4. Use several GPUs, a large unified-memory machine, or CPU offload (slow).

Can a Mac run a 500B–800B model?

None of the Macs or mini PCs we track fit it at Q4.

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 500B–800B model?

Short answer: about 478.2 GB at Q4_K_M, the usual quant for local use. Below: every quant, longer contexts, the cheapest hardware that fits, and every 500B–800B open model we track.

Q4_K_M: 478.2 GB 415.0 GB weights
Q8_0: 837.4 GB 727.3 GB weights
FP16: 1575.3 GB 1369.0 GB weights

Worked out for DeepSeek-R1 (684.5B parameters, from the Hugging Face API). Formula: weights × 1.15 + 1 GB at a ~4k context (how we calculate).

No single consumer GPU we track fits it at Q4.

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 500B–800B model at each quant
Setting	Memory needed	Cheapest GPU	Cheapest Mac / mini PC
Q4_K_M, ~4k context	478.2 GB	None we track	None we track
Q8_0, ~4k context	837.4 GB	None we track	None we track
FP16, ~4k context	1575.3 GB	None we track	None we track

No single consumer GPU we track holds a 500B–800B model at Q4. Your options: two or more GPUs (see the ready-made builds or the planner), a large unified-memory machine (see the best Mac for local LLMs), or a smaller quant with lower quality.

Hardware that runs it

Bar: Q4 memory needed vs. usable memory. Speed is estimated decode at Q4 (how we estimate); MoE shows an upper bound only.FitsTight

None of the devices we track run it at Q4.

500B–800B models we track

Memory needed by each 500B–800B model we track
Model	Q4_K_M	Q8_0	FP16
Beam-1.0-501B-A23B501.3B params · reflection · MoE	350.5 GB	613.5 GB	1153.9 GB
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16560.5B params · nvidia · MoE	391.8 GB	685.9 GB	1290.2 GB
LongCat-Flash-Chat561.9B params · meituan-longcat	392.7 GB	687.5 GB	1293.3 GB
LongCat-Flash-Thinking561.9B params · meituan-longcat	392.7 GB	687.5 GB	1293.3 GB
LongCat-Flash-Thinking-2601561.9B params · meituan-longcat	392.7 GB	687.5 GB	1293.3 GB
Ling-3.1-flash572.1B params · inclusionAI · MoE	399.9 GB	700.1 GB	1316.9 GB
Ling-3.1-flash-1019572.1B params · inclusionAI · MoE	399.9 GB	700.1 GB	1316.9 GB
Step-5-Preview604.3B params · stepfun-ai · MoE	422.3 GB	739.4 GB	1391.0 GB
DeepSeek-R1684.5B params · deepseek-ai · MoE	478.2 GB	837.4 GB	1575.3 GB
DeepSeek-R1-Zero684.5B params · deepseek-ai · MoE	478.2 GB	837.4 GB	1575.3 GB
DeepSeek-V3-0324684.5B params · deepseek-ai · MoE	478.2 GB	837.4 GB	1575.4 GB
DeepSeek-R1-0528684.5B params · deepseek-ai · MoE	478.2 GB	837.4 GB	1575.4 GB
DeepSeek-V3.1-Base684.5B params · deepseek-ai · MoE	478.2 GB	837.4 GB	1575.4 GB
DeepSeek-V3.2685.4B params · deepseek-ai · MoE	478.8 GB	838.4 GB	1577.3 GB
DeepSeek-V3.2-Exp685.4B params · deepseek-ai · MoE	478.9 GB	838.5 GB	1577.4 GB
DeepSeek-V3.2-Speciale685.4B params · deepseek-ai · MoE	478.9 GB	838.5 GB	1577.4 GB
K-EXAONE-2.0-750B-A37B749.4B params · LGAI-EXAONE · MoE	523.4 GB	916.6 GB	1724.5 GB
GLM-5.3753.3B params · zai-org	526.2 GB	921.5 GB	1733.7 GB
GLM-5.2753.3B params · zai-org · MoE	526.2 GB	921.5 GB	1733.7 GB
GLM-5.3-BF16753.3B params · zai-org · MoE	526.2 GB	921.5 GB	1733.7 GB
GLM-5753.9B params · zai-org · MoE	526.6 GB	922.1 GB	1734.9 GB
GLM-5.1753.9B params · zai-org · MoE	526.6 GB	922.1 GB	1734.9 GB
Hy4-preview780.0B params · tencent · MoE	544.8 GB	954.0 GB	1794.9 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 500B–800B model?: About 478.2 GB at Q4_K_M, 837.4 GB at Q8_0 and 1575.3 GB at FP16, at a ~4k context. That is for DeepSeek-R1 (684.5B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.
What is the cheapest GPU for a 500B–800B model?: None of the single consumer GPUs we track hold it at Q4. Use several GPUs, a large unified-memory machine, or CPU offload (slow).
Can a Mac run a 500B–800B model?: None of the Macs or mini PCs we track fit it at Q4.
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 · 30B–32B · 40B–50B · 70B · 120B · 200B–300B · 300B–480B · 1T+ · All sizes · Ollama VRAM by tag · GGUF sizes by quant

How much VRAM do you need for a 500B–800B model?

Cheapest hardware for each quant

Hardware that runs it

500B–800B models we track

Quick answers

New models, and what they need to run