Can I run DeepSeek V4 Pro locally?
DeepSeek V4 Pro by deepseek-ai has 1.60T parameters (51B active per token, 384 experts, 6 used per token). At 4-bit with an 8K context it needs about 891 GB of GPU memory. It fits on 0 of 65 GPUs and Macs we track.
This model uses a newer compressed or sparse attention design. Our KV-cache numbers use the standard formula, so treat long-context figures as an upper bound.
Memory needed by quant and context length
GB of GPU memory = weights + KV cache (FP16) + runtime overhead. Weights use real file sizes from teamblobfish/DeepSeek-V4-Pro-GGUF where available (marked •), otherwise parameter count × bits per weight. How we calculate.
| Quant | Quality | 4K ctx | 8K ctx | 32K ctx | 128K ctx |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | rough | 536 GB | 536 GB | 540 GB | 553 GB |
| 3-bit (Q3_K_M) | usable | 729 GB | 729 GB | 733 GB | 746 GB |
| 4-bit (Q4_K_M) • | good, most popular | 890 GB | 891 GB | 894 GB | 907 GB |
| 5-bit (Q5_K_M) | very good | 1062 GB | 1063 GB | 1066 GB | 1079 GB |
| 6-bit (Q6_K) | near lossless | 1222 GB | 1223 GB | 1226 GB | 1239 GB |
| 8-bit (Q8_0) • | effectively lossless | 1558 GB | 1559 GB | 1562 GB | 1575 GB |
| 16-bit (F16/BF16) | original weights | 2979 GB | 2980 GB | 2983 GB | 2996 GB |
Which GPUs and Macs run DeepSeek V4 Pro?
Best quality quant that fits at 8K context. "Tight" = within 10% of usable memory. Speed is a ceiling from memory bandwidth (real speed is typically 50–80% of it).
Prefer an API?
DeepSeek: DeepSeek V4 Pro 0423 is on OpenRouter at $0.90 per 1M input tokens and $1.50 per 1M output tokens. Price details and history →
Questions
How much VRAM does DeepSeek V4 Pro need?
About 891 GB at 4-bit (Q4_K_M) with an 8K context: 889 GB of weights (real GGUF file size), 1.1 GB of KV cache and 0.6 GB of runtime overhead. At 8-bit it needs about 1559 GB.
What is the smallest GPU that runs DeepSeek V4 Pro?
No single consumer GPU in our table fits it at Q4_K_M with 8K context.
Can I run DeepSeek V4 Pro on a Mac?
Not at Q4_K_M with 8K context on any Mac in our table, even a 512 GB Mac Studio, using the default GPU memory limit.
Model facts
| Hugging Face repo | deepseek-ai/DeepSeek-V4-Pro |
| Parameters | 1,598,839,674,782 (1.60T) · ~51B active (computed from config) |
| Layers / hidden size | 61 / 7168 |
| Attention | GQA; 128 heads, 1 KV heads, head dim 512 |
| KV cache per token (FP16) | 137.3 KiB |
| Max context | 1,048,576 tokens |
| License | mit |
| Released on HF | Apr 22, 2026 |
| Downloads (30d) / likes | 311,047 / 5,717 |
Sources: Hugging Face model API · config.json (deepseek-ai/DeepSeek-V4-Pro) · GGUF files · GPU specs from vendor pages (list). Data fetched Nov 17, 2026. Spot an error? Report it.