Can you run DeepSeek V4 Pro DSpark locally?
DeepSeek V4 family · Jun 2026 (DSpark attach) · 1.6T parameters (49B activated per token) · Hugging Face model card
DeepSeek V4-Pro-DSpark is not a new model (official README): it is the Pro checkpoint with a speculative-decoding module attached. Config architecture matches Pro (deepseek-ai/DeepSeek-V4-Pro-DSpark@7c09739fd136abfb70a49ec334157f65f45b52cd): 61 layers, 1.6T/49B, 384 experts × 6, 1M context, MIT. safetensors.total=1,650,497,936,906 recorded as dump size only. Measured tok/s: not published here.
Minimum: Same as V4-Pro: multi-node / ~1056 GB at Q4_K_M (calculated from 1.6T resident) · Recommended: Same as V4-Pro plus DSpark speculative decoding enabled at serve time
How much VRAM does DeepSeek V4 Pro DSpark need at each quantization?
DeepSeek V4 Pro DSpark needs 1,056.5 GB of VRAM at Q4_K_M. The table below lists weights-only size and total VRAM including overhead for each common quantization level with 4k and 32k token contexts.
| Quantization | Bits / weight | Weights only | Total + KV @4k ctx | Total + KV @32k ctx |
|---|---|---|---|---|
| Q4_K_M | 4.8 | 960.0 GB | 1,056.5 GB | 1,060.1 GB |
| Q5_K_M | 5.7 | 1,140.0 GB | 1,254.5 GB | 1,258.1 GB |
| Q6_K | 6.6 | 1,320.0 GB | 1,452.5 GB | 1,456.1 GB |
| Q8_0 | 8.5 | 1,700.0 GB | 1,870.5 GB | 1,874.1 GB |
| FP16 | 16 | 3,200.0 GB | 3,520.5 GB | 3,524.1 GB |
Official: not a new model. Same Pro arch keys. VRAM follows 1.6T resident.
Method: weights = parameters × bits-per-weight (Q4_K_M ≈ 4.8, Q5_K_M ≈ 5.7, Q6_K ≈ 6.6, Q8_0 ≈ 8.5, FP16 = 16), plus 10% loading overhead, plus KV cache from the verified architecture config (61 layers, 1 KV heads, 512 head dim). Source: HF-verified 2026-09-21 via api (gated=false private=false safetensors.total=1650497936906) + config.json repo=deepseek-ai/DeepSeek-V4-Pro-DSpark@7c09739fd136abfb70a49ec334157f65f45b52cd model_type=deepseek_v4 layers=61 hidden=7168 heads=128 kv_heads=1 head_dim=512 ctx=1048576 n_routed_experts=384 num_experts_per_tok=6 n_shared_experts=1 sliding_window=128 license=mit. Vendor README table: total=1600B active=49B precision=fp4_fp8_mixed. VRAM GB are T08 documented calculations from resident total, not measurements. No tok/s invented. Official: DSpark is not a new model.
Which GPUs can run DeepSeek V4 Pro DSpark locally?
At Q4_K_M with a 4k context, DeepSeek V4 Pro DSpark needs 1,056.5 GB of VRAM. The lists below are computed live from our GPU database and grouped by how much headroom the card has. Disclosure: CompareAIHardware.com participates in the Amazon Associates program and earns from qualifying purchases through links on this page. Affiliate relationships do not influence our recommendations.
No single consumer GPU in our database runs DeepSeek V4 Pro DSpark at Q4_K_M. The 1,056.5 GB requirement exceeds every listed consumer card; see the Mac and data center options above and below.