DeepSeek V4-Flash — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
284B-total MoE with 13B active parameters per token, 1M context, MIT license. Even at Q4 this needs roughly 140-160GB depending on quant — realistically a 2-4x RTX 4090/5090 workstation with heavy CPU offload, or a single 80GB+ datacenter GPU. Local inference is experimental: only community llama.cpp forks (e.g. antirez's branch, ~18 tok/s on a 96GB RTX 6000) run it today; no mainline llama.cpp or stable local Ollama build yet. Treat as 'enthusiast workstation, WIP software' rather than a turnkey local model.
DeepSeek V4-Flash needs about 172 GB of VRAM at Q4 (experimental) — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Weights plus overhead only: this model's architecture is unpublished, so no KV cache is added.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 94.2 GB | — | Too big | |
| Q3_K_M 3.41 bpw | 121.9 GB | — | Too big | |
| Q4_K_M 4.83 bpw | 172.3 GB | — | Too big | |
| Q5_K_M 5.67 bpw | 202.1 GB | — | Too big | |
| Q6_K 6.56 bpw | 233.7 GB | — | Too big | |
| Q8_0 8.50 bpw | 302.6 GB | — | Too big | |
| F16 16.00 bpw | 568.8 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. DeepSeek V4-Flash VRAM calculator →
Get DeepSeek V4-Flash running
The cheapest catalogued GPU that runs DeepSeek V4-Flash is the Apple M3 Ultra (512 GB).
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run DeepSeek V4-Flash
Cloud-hosted on Ollama. Local inference is work-in-progress llama.cpp forks only.
Specifications
Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.
Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.
- Parameters
- 284 Billion (13B active)
- Context window
- 1,000,000
- Architecture
- Mixture-of-Experts
- Provider
- DeepSeek
- Licence
- MIT
- Practical deployment
- Workstation (large unified memory or pro GPU)
- Specified at
- Q4 (experimental)
- System RAM
- 256 GB
- Record updated
- 2026-04-24
Commercial use permitted. No usage restrictions beyond attribution.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Can I run DeepSeek V4-Flash on my GPU?
- DeepSeek V4-Flash on AMD Ryzen AI Max+ 395
- DeepSeek V4-Flash on Apple M1 Ultra
- DeepSeek V4-Flash on Apple M2 Max
- DeepSeek V4-Flash on Apple M2 Ultra
- DeepSeek V4-Flash on Apple M4 Max
- DeepSeek V4-Flash on Apple M5 Max
- DeepSeek V4-Flash on NVIDIA A100 80GB (PCIe)
- DeepSeek V4-Flash on NVIDIA DGX Spark
- DeepSeek V4-Flash on NVIDIA H100 80GB (PCIe)
- DeepSeek V4-Flash on NVIDIA RTX PRO 6000 Blackwell
Other DeepSeek V4 sizes
DeepSeek V4-Flash — frequently asked questions
How much VRAM does DeepSeek V4-Flash need?
About 172 GB at Q4 (experimental) — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does DeepSeek V4-Flash run on an RTX 4090 (24 GB)?
No. DeepSeek V4-Flash needs about 172 GB at Q4 (experimental), more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run DeepSeek V4-Flash locally?
Cloud-hosted on Ollama. Local inference is work-in-progress llama.cpp forks only. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does DeepSeek V4 come in?
DeepSeek V4-Flash (172 GB), DeepSeek V4-Pro (967 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.