How to Run DeepSeek V4 Flash Locally: GGUF, Hardware, and Setup
DeepSeek shipped the official V4 Flash build today (July 31, 2026, build name DeepSeek-V4-Flash-0731), and unlike most launch days, this one comes with everything a local user wants: open weights on Hugging Face under MIT, GGUF quantizations on day one, and a model shape that real hardware can hold. Here is how to run it on your own machine, from the 5 GB shortcut to the full 155 GB build. (New to the release itself? Start with our V4 Flash 0731 explainer. Prefer someone else's GPU? See the cloud guide.)
Why This One Is Different
V4 Flash is a 284B parameter Mixture of Experts model that activates only 13B parameters per token. That ratio is the whole story for local use: you need enough memory to hold the weights, but the compute per token is small, so CPU plus GPU setups reach usable speeds instead of crawling. Compare that with the 1.6T V4 Pro (server territory, see the V4 Pro guide) or the 2.8T Kimi K3 (not happening at home), and V4 Flash lands in a rare spot: frontier grade agent benchmarks in a package a maxed out workstation or Mac Studio can run.
Your Options at a Glance
| Option | Download size | Hardware class | Best for |
|---|---|---|---|
| Qwen 3.5 9B V4 Distill | 5.3 GB | Any modern PC (8 GB RAM / 6 GB VRAM) | Everyone: the V4 reasoning style in a small file |
| V4 Flash, 4 bit GGUF | 144 GB (preview) / 155 GB (0731) | 192 GB+ memory (Mac Studio, big workstation) | The real thing, fully local |
| V4 Flash Abliterated | 154 GB | Same as above | The uncensored build (details) |
| Cloud (LU Labs, official API) | none | Any device | No 150 GB download, see the cloud guide |
What You Need for the Full Model
The unsloth GGUFs of the 0731 build come in two flavors: UD-Q4_K_XL at 155 GB and UD-Q8_K_XL at 162 GB. The 8 bit file is only 7 GB larger than the 4 bit one (MoE expert weights dominate and quantize well), so if you have the memory, take Q8. On top of the weights, budget headroom for the KV cache and your OS. Realistic targets:
- Mac Studio with 192 GB unified memory: runs the 4 bit build with room for context. The most comfortable single box for this model.
- Workstation with 192 to 256 GB DDR5 plus a 24 GB GPU: put attention and shared layers on the GPU, experts in RAM. Thanks to the 13B active parameters, expect interactive speeds, roughly in the 5 to 15 tokens per second range depending on memory bandwidth.
- Anything smaller: take the distill (next section) or the cloud route.
We ran the full memory math, including context length costs, in Can You Run DeepSeek V4 Flash Locally?
The 5 Minute Path: One Click in a Local AI Studio
Locally Uncensored is a free, open source local AI studio: one installer gives you chat, image and video generation, and a coding agent on your own PC, with no command line and no Docker. The V4 Flash family is already in the built in model catalog:
- Install Locally Uncensored and let the setup wizard detect or install a backend.
- Open the Models tab and search for DeepSeek.
- Pick your size: Qwen 3.5 9B DeepSeek V4 Distill (5.3 GB), DeepSeek V4 Flash (the full 144 GB preview build, multi part download), or DeepSeek V4 Flash Abliterated (154 GB).
- Click download, wait, chat. The app checks your VRAM and warns you before you grab something your machine cannot hold.
The 0731 Build with llama.cpp
The in app catalog entry currently points at the preview weights; the fresh 0731 GGUFs landed on Hugging Face today at unsloth/DeepSeek-V4-Flash-0731-GGUF. With a current llama.cpp, one line pulls and serves it:
llama serve -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-Q4_K_XLThat exposes an OpenAI compatible server on localhost. Point Locally Uncensored (or any client) at it: the app auto detects llama.cpp among its 12 supported backends, so the model shows up in the picker like any other. If you manage downloads yourself, you can also set a custom HuggingFace folder in Settings and let the app index the file from there.
The Small Path: Qwen 3.5 9B V4 Distill
Not everyone owns 192 GB of RAM, and the community answer arrived fast: Qwen3.5-9B-DeepSeek-V4-Flash distills the V4 Flash reasoning behavior into a 9B model. It is a 5.3 GB download at Q4_K_M, runs on a 6 GB GPU or plain CPU, and gives you a real taste of the V4 style: structured reasoning, strong tool use for its size, 1M context lineage. It sits in the same in app catalog, one click away. It is a distill of the preview checkpoint; expect an 0731 refresh from the community soon.
Uncensored: The Abliterated Build
The huihui abliterated version of V4 Flash is the most downloaded uncensored model of 2026 so far (over 600K pulls). Same architecture, same hardware needs, refusal behavior removed. We cover what abliteration does and does not change in the dedicated post, and the background in our abliterated models guide.
FAQ
Is it free for commercial use? Yes, MIT license, both preview and 0731.
Does the 1M context work locally? The architecture supports it, but KV cache memory makes very long contexts expensive at home. 32K to 128K is the practical local range; go cloud for the full window.
Windows or Linux? Both. Locally Uncensored ships installers for Windows 10/11 and Linux (deb, rpm, AppImage).