Can you run DeepSeek V4.1 Flash locally?
deepseek family · Sep 2026 · 552B parameters (16B activated per token) · Hugging Face model card
Multimodal MoE with a 552B backbone. Causal Encoder-Decoder activates about 8B parameters per token in prefill and 16B in decode, with a 1M-token context and MIT weights. Resident Q4-class weights are still hundreds of gigabytes, so the floor is dual DGX Spark or 256 GB unified memory; 128 GB machines need SSD expert streaming. Measured local tok/s for this 2026-09-10 architecture: not yet published here.
Minimum: 2x DGX Spark or 256 GB unified memory (Q4-class / expert offload) · Recommended: 256 GB+ Mac Studio or multi-GPU server for serving
How much VRAM does DeepSeek V4.1 Flash need at each quantization?
DeepSeek V4.1 Flash needs 364.3 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 | 331.2 GB | 364.3 GB | 364.3 GB |
| Q5_K_M | 5.7 | 393.3 GB | 432.6 GB | 432.6 GB |
| Q6_K | 6.6 | 455.4 GB | 500.9 GB | 500.9 GB |
| Q8_0 | 8.5 | 586.5 GB | 645.2 GB | 645.2 GB |
| FP16 | 16 | 1,104.0 GB | 1,214.4 GB | 1,214.4 GB |
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 (not applicable to this architecture). Source: DeepSeek-V4.1-Flash official card 2026-09-10: 552B backbone MoE, Causal Encoder-Decoder (40 layers = 20 encoder + 20 decoder), 8B active/token at prefill and 16B at decode, 1M context, MIT, multimodal image-text-to-text. CSA2 + FP4 main KV (~890 B/token). Engram conditional memory 196B sparsely accessed (not in the 552B backbone count). Sources: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash README + config.json (text_config.max_position_embeddings=1048576, n_routed_experts=384, num_experts_per_tok=6). Prior 284B/13B '0731 refresh' figures described DeepSeek-V4-Flash, not this repo. Safetensors dump can exceed 552B because the repo ships multiple precision copies. Cycle 2026-09-17.
Which GPUs can run DeepSeek V4.1 Flash locally?
At Q4_K_M with a 4k context, DeepSeek V4.1 Flash needs 364.3 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.
Minimum GPUs that fit DeepSeek V4.1 Flash at Q4_K_M
These GPUs hold the model but leave little headroom — keep contexts short.
| GPU | VRAM | Bandwidth | Type | |
|---|---|---|---|---|
| AMD Instinct MI430X | 432 GB | 23300 GB/s | Data Center GPU | |
| AMD Instinct MI455X | 432 GB | 23300 GB/s | Data Center GPU |
Needs 2+ GPUs to run DeepSeek V4.1 Flash
One of these cards is too small on its own, but a pair (tensor or pipeline parallel, ~90% efficiency) covers the 364.3 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 MI350X | 288 GB ×2 | 8000 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 |
Can a Mac run DeepSeek V4.1 Flash?
Yes — these Apple Silicon machines fit DeepSeek V4.1 Flash at Q4_K_M, since macOS lets the GPU use about 75% of unified memory. Generation speed is bound by memory bandwidth, so the GB/s column matters as much as capacity.
| Apple Silicon | Unified memory | Usable by GPU (~75%) | Bandwidth |
|---|---|---|---|
| M5 Ultra (Mac Studio) | 512 GB | 384 GB | 1200 GB/s |