Can you run DeepSeek V4 Flash locally?
DeepSeek V4 family · Apr 2026 (preview) · 284B parameters (13B activated per token) · Hugging Face model card
DeepSeek V4-Flash is a mixture-of-experts model with 284B total parameters and 13B activated per token (official README table; MIT). Config (deepseek-ai/DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1) is model_type=deepseek_v4, 43 layers, hidden_size=4096, 64 attention heads, 1 KV head, head_dim=512, 256 routed experts × 6 per token + 1 shared, native context 1,048,576, sliding_window=128, YaRN rope_scaling factor 16. Shipped precision is FP4+FP8 mixed (experts FP4, most other params FP8). Resident Q4_K_M weights need about 187 GB (T08 documented calculation from 284B total — all experts must sit in memory; do not use the 13B active count as VRAM). No consumer GPU holds it. A 256 GB Mac Studio has 256 GB unified_memory_max; macOS GPU usable share is about 75% (192 GB, T13), so Q4_K_M is a tight/over-budget fit on usable-not-raw memory. Measured tok/s: not published here.
Minimum: 1× NVIDIA B200 192GB (Q4_K_M ~187 GB, calculated from 284B resident) or 4× 80GB · Recommended: 2× NVIDIA B200 192GB or 8× H100 80GB; 256 GB unified memory is usable-not-raw (see T13 75% macOS GPU share)
How much VRAM does DeepSeek V4 Flash need at each quantization?
DeepSeek V4 Flash needs 187.8 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 | 170.4 GB | 187.8 GB | 190.3 GB |
| Q5_K_M | 5.7 | 202.4 GB | 223.0 GB | 225.5 GB |
| Q6_K | 6.6 | 234.3 GB | 258.1 GB | 260.6 GB |
| Q8_0 | 8.5 | 301.8 GB | 332.3 GB | 334.8 GB |
| FP16 | 16 | 568.0 GB | 625.2 GB | 627.7 GB |
DeepSeek V4 hybrid CSA+HCA (official README). kv_heads=1, head_dim=512, sliding_window=128. VRAM floor follows 284B resident weights, not 13B active.
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 (43 layers, 1 KV heads, 512 head dim). Source: HF-verified 2026-09-21 via api (gated=false private=false safetensors.total=290944616402) + config.json repo=deepseek-ai/DeepSeek-V4-Flash@60d8d70770c6776ff598c94bb586a859a38244f1 model_type=deepseek_v4 layers=43 hidden=4096 heads=64 kv_heads=1 head_dim=512 ctx=1048576 n_routed_experts=256 num_experts_per_tok=6 n_shared_experts=1 sliding_window=128 license=mit. Vendor README table: total=284B active=13B precision=fp4_fp8_mixed. VRAM GB are T08 documented calculations from resident total, not measurements. No tok/s invented.
Which GPUs can run DeepSeek V4 Flash locally?
At Q4_K_M with a 4k context, DeepSeek V4 Flash needs 187.8 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.
Runs DeepSeek V4 Flash comfortably (25%+ VRAM headroom)
| GPU | VRAM | Bandwidth | Type | |
|---|---|---|---|---|
| AMD Instinct MI325X | 256 GB | 6000 GB/s | Data Center GPU | |
| AMD Instinct MI350X | 288 GB | 8000 GB/s | Data Center GPU | |
| AMD Instinct MI355X | 288 GB | 8000 GB/s | Data Center GPU | |
| NVIDIA B300 (Blackwell Ultra) | 288 GB | 8000 GB/s | Data Center GPU | |
| AMD Instinct MI430X | 432 GB | 23300 GB/s | Data Center GPU | |
| AMD Instinct MI455X | 432 GB | 23300 GB/s | Data Center GPU |
Minimum GPUs that fit DeepSeek V4 Flash at Q4_K_M
These GPUs hold the model but leave little headroom — keep contexts short.
| GPU | VRAM | Bandwidth | Type | |
|---|---|---|---|---|
| AMD Instinct MI300X | 192 GB | 5300 GB/s | Data Center GPU | |
| NVIDIA B200 | 192 GB | 8000 GB/s | Data Center GPU |
Needs 2+ GPUs to run DeepSeek V4 Flash
One of these cards is too small on its own, but a pair (tensor or pipeline parallel, ~90% efficiency) covers the 187.8 GB requirement. See our multi-GPU guide for setup.
| GPU | VRAM | Bandwidth | Type | |
|---|---|---|---|---|
| H200 SXM | 141 GB ×2 | 4800 GB/s | Data Center GPU |
Can a Mac run DeepSeek V4 Flash?
Yes — these Apple Silicon machines fit DeepSeek V4 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 |
Which pre-built systems can run DeepSeek V4 Flash?
These mini PCs and workstations from our database fit DeepSeek V4 Flash at Q4_K_M. Usable-memory figures are conservative: Windows shares about half of system RAM with the GPU by default (Linux can expose more), macOS lets Apple Silicon GPUs use about 75% of unified memory, and Linux unified-memory systems such as GB10 expose roughly 90%. Generation speed is bound by memory bandwidth, so compare the GB/s column before buying.
| System | Type | Memory | Usable for AI | Bandwidth | GPUs |
|---|---|---|---|---|---|
| HP Z8 Fury G5 (4x RTX 6000 Ada) | Workstation | 192 GB | 192 GB | 960 GB/s | 4 |
Frequently asked questions
Can you run DeepSeek V4-Flash locally?
Only on data-center or large unified-memory hardware. Resident Q4_K_M weights need about 187 GB (calculated from the official 284B total). That is a single 192 GB NVIDIA B200 or at least four 80 GB GPUs. No consumer graphics card can run it without heavy offloading.
Why is VRAM so large if only 13B parameters activate per token?
DeepSeek V4-Flash is a Mixture-of-Experts model: all 284 billion parameters must sit in memory even though only 13 billion are active per token, according to the official Hugging Face README. VRAM use follows total resident parameters; generation speed follows activated parameters. Measured tok/s are not published here.
Can a Mac Studio run DeepSeek V4-Flash?
A Mac Studio with M3 Ultra and 256 GB unified_memory_max (product spec) exposes about 192 GB to the GPU under the T13 macOS 75% usable-memory rule. The Q4_K_M resident estimate is about 187 GB, so capacity is a tight usable-not-raw fit, not a raw-256-GB fit. Generation speed is limited by 819 GB/s memory bandwidth. Measured tok/s: not published here.
Related guides
- Open DeepSeek V4 Flash in the VRAM calculator — pick quantization and context, see the memory math per GPU class.
- DeepSeek V4 Flash 0731 Guide
- Best GPUs for LLMs
- Multi-GPU Guide