How much VRAM does DeepSeek V4 Pro 1.6T need at Q4_K_M? CSA + MLA frontier
Resident-mode V4 Pro at 1M context is datacenter territory: ~13 H100 80GB (two 8-GPU nodes) or ~8 H200 141GB with NVLink, or comparable MI300X capacity. For most workloads the hosted DeepSeek API is dramatically cheaper than self-hosting at this scale - self-host only when data residency or bit-exact control is the real requirement.
DeepSeek V4 Pro 1.6T at Q4_K_M with the full 1M-token context needs about 997 GB of VRAM with every expert loaded. That's the real shape of the model and the number to plan a deployment against. The 1.6T parameter pool at Q4 is 896 GB of weights on its own. Compressed attention keeps the KV cache to about 10.3 GB even at 1M tokens, and runtime overhead adds about 91 GB. Each token only uses 49B parameters, but that doesn't shrink what has to stay loaded, because different tokens pick different experts. Moving experts to system RAM cuts VRAM, but the host then has to hold those weights and every token gets much slower.
By TechCompare · Updated
Calculator
Estimated VRAM required
997 GB
1600B params at Q4_K_M, 1,048,576 token context, batch 1, inference.
Estimate accuracy: Weights within ~2%. KV cache within ~5% for standard GQA models, ~10% for MLA or hybrid state estimates. Real VRAM may vary with framework (vLLM vs llama.cpp vs Transformers), Flash Attention, and driver overhead.
Compressed attention applied: This model stores far fewer cache entries than one per token, so the KV cache shown is 8% of what the standard per-token formula gives for these settings.
Hardware that fits
Apple entries show installed unified memory separately from a model budget based on an 8 GB reserve. That budget is a planning heuristic, not a fixed hardware limit.
No single GPU in our catalog has enough memory. Multi-GPU or CPU offload required.
How this is calculated
The 1.6T parameter pool dominates: 1,600B x 0.56 bytes per parameter = 896 GB of weights, no matter which experts a given token routes to. The KV cache follows DeepSeek's reference code, not the standard per-token formula. K and V share one 512-value vector per cache entry. 30 of the 61 layers use Compressed Sparse Attention (CSA) and keep one entry per 4 tokens, plus a 128-value indexer key. The other 31 use Heavily Compressed Attention (HCA) and keep one entry per 128 tokens. That averages about 80.7 values per token per layer, or 10.3 GB in 16-bit at the full 1,048,576-token window. Each layer also keeps its last 128 tokens uncompressed, but that's a fixed 8 MB or so. The calculator adds 10% of weights plus cache as overhead, about 90.6 GB, for 997 GB total.
Verdict
The budget comes down to three numbers: 896 GB of weights at Q4 (the full 1.6T pool times the 0.56 B/param rate), about 10.3 GB of compressed KV cache at the full 1M window, and about 91 GB of runtime overhead, totaling 997 GB resident. The 49B active count doesn't lower that total. Any token can route to any expert, so every expert stays in VRAM for full-speed serving. CSA and HCA are what keep the long-context cache this small. Without that compression, one 512-value entry per token in all 61 layers would need about 65.5 GB at 1M tokens, over six times the cache shown here. Plan on resident mode for serving.
More DeepSeek scenarios
Related guides
Frequently asked questions
What's CSA and why does it matter?
Can V4 Pro fit on 8x H100?
Should I run V4 Pro or V4 Flash locally?
Related tools
RAM Latency Calculator
Convert DDR3/DDR4/DDR5 timings (CL, tRCD, tRP, tRAS) into true latency in nanoseconds.
Use tool ➜Power Cost Estimator
Estimate annual electricity costs for your PC, Server, or TV.
Use tool ➜Data Transfer Calculator
Estimate transfer times for files over USB, WiFi, Ethernet, and more.
Use tool ➜Memory and Storage Throughput Visualizer
Visualize the massive speed difference between CPU cache, RAM, and storage.
Use tool ➜