Can you run DeepSeek-V3.2 locally?
DeepSeek V3.2 family · 2025 · 685.4B parameters (37B activated per token) · Hugging Face model card
DeepSeek-V3.2 is a mixture-of-experts open-weight model with roughly 685B total parameters and 37B active per token (config: 256 routed experts, 8 activated per token, 1 shared expert; HF safetensors.total=685,355,329,792). Its headline change over V3 is DeepSeek Sparse Attention (DSA), which cuts long-context compute while keeping quality. Native context is 163,840 tokens (config max_position_embeddings), MIT licensed. Local VRAM fit uses the 37B active count, not the 685B dense-equivalent: roughly 24 GB at Q4-class (calculated); full weights still occupy ~1.4 TB at bf16, so real deployments run multi-GPU. Measured local speeds: not yet published here.
Minimum: 24 GB+ memory (Q4-class on 37B active, calculated) · Recommended: 32 GB+ for comfortable context headroom (Q4-class on 37B active, calculated)
How much VRAM does DeepSeek-V3.2 need at each quantization?
DeepSeek-V3.2 needs 452.3 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 | 411.2 GB | 452.3 GB | 452.3 GB |
| Q5_K_M | 5.7 | 488.3 GB | 537.2 GB | 537.2 GB |
| Q6_K | 6.6 | 565.4 GB | 622.0 GB | 622.0 GB |
| Q8_0 | 8.5 | 728.2 GB | 801.0 GB | 801.0 GB |
| FP16 | 16 | 1,370.7 GB | 1,507.8 GB | 1,507.8 GB |
MLA qk_nope=128 qk_rope=64 v=128; DSA sparse attention (yarn factor=40)
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, 128 KV heads, head dim). Source: HF-verified 2026-10-06 via api (gated=false private=false params=685.36B safetensors.total=685355329792 ctx=163840 license=mit repo=deepseek-ai/DeepSeek-V3.2@a7e62ac04ecb2c0a54d736dc46601c5606cf10a6 layers=61 hidden=7168 heads=128 kv_heads=128 experts=256 per_tok=8). active_params_b=37.0 (DeepSeek V3.2 lineage active count (37B routed+shared active, same MoE budget as V3/R1 architecture; config n_routed_experts=256, n_experts_per_tok=8)). VRAM figures are documented calculations (Q4-class, 0.6 GB per B params), not measurements. No tok/s invented.
Which GPUs can run DeepSeek-V3.2 locally?
At Q4_K_M with a 4k context, DeepSeek-V3.2 needs 452.3 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.
Needs 2+ GPUs to run DeepSeek-V3.2
One of these cards is too small on its own, but a pair (tensor or pipeline parallel, ~90% efficiency) covers the 452.3 GB requirement. See our multi-GPU guide for setup.
| GPU | VRAM | Bandwidth | Type | |
|---|---|---|---|---|
| AMD Instinct MI325X | 256 GB ×2 | 6000 GB/s | Data Center GPU | |
| AMD Instinct MI355X | 288 GB ×2 | 8000 GB/s | Data Center GPU | |
| NVIDIA B300 (Blackwell Ultra) | 288 GB ×2 | 8000 GB/s | Data Center GPU |