How to Download DeepSeek V4: HuggingFace & Ollama

What's up, self-hosters — if you searched "how to download DeepSeek V4" hoping for a straight answer instead of a pile of speculation, you're in the right place.

I've been running DeepSeek's open-weight releases locally since V3, and V4 isn't a rumor anymore. DeepSeek shipped it as an open-weight preview on April 24, 2026: two MoE checkpoints, DeepSeek-V4-Pro and DeepSeek-V4-Flash, both sitting on Hugging Face under the verified deepseek-ai organization. That part's simple. What's not simple is everything else you'll run into the moment you search for a shortcut.

Before You Download Anything

Here's the part most "how to download" guides skip entirely.

Within about 45 minutes of DeepSeek's official V4 announcement, someone spun up a fake GitHub organization named DeepSeek-V4, complete with the real DeepSeek logo and genuine benchmark charts copied straight from the launch post, then attached two installer archives to a release. Anyone who downloaded and ran them got the Vidar infostealer instead of a language model. Microsoft's threat intelligence team tracked the entire timeline — fake org live within 45 minutes, first confirmed victim about four hours later, 91 stars and 27 forks accumulated within four days. For a stretch after that, the fake repo actually outranked the real Hugging Face page on Bing for searches like "DeepSeek v4 weights github." Full writeup is on the official Microsoft Security Blog, if you want the technical detail.

The fact worth remembering: DeepSeek has never published the V4 weights on GitHub. DeepSeek's real GitHub organization, github.com/deepseek-ai, is legitimate and hosts code and tooling — it's just not where the model itself lives. If a download path routes through GitHub, a Discord or Telegram invite, a "mirror" someone posted in a forum, or anywhere other than huggingface.co/deepseek-ai, chat.deepseek.com, or api.deepseek.com, it isn't DeepSeek.

A second tell, worth keeping in your back pocket: DeepSeek-V4-Pro alone is a 1.6-trillion-parameter model. A real download never lands as a tidy 100MB installer.exe. That was the giveaway in the case Microsoft documented — the entire fake "model" was two 102MB archives wrapped around a malware loader. If what you're looking at fits on a USB stick, it's not the model.

3 Ways to Access DeepSeek V4

If you're not sure which one you need, start with the API or chat.deepseek.com. It costs nothing to try and tells you in five minutes whether the multi-hundred-gigabyte download is actually worth it for what you're doing.

Option 1 — Official API & Web Chat (No Download)

This is the lowest-risk way to use V4, and honestly the one most people actually want. DeepSeek's own interface at chat.deepseek.com runs both models through Expert Mode and Instant Mode, per DeepSeek's launch announcement, with no install step at all.

For programmatic access, the API is OpenAI-compatible and now also speaks the Anthropic Messages format natively, so dropping it into an existing integration is mostly a base URL and model-name swap:

from openai import OpenAI
client = OpenAI(
    api_key="<your DeepSeek API key>",
    base_url="https://api.deepseek.com",
)
response = client.chat.completions.create(
    model="deepseek-v4-flash",  # or "deepseek-v4-pro"
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"}
    ],
    temperature=1.0,
    top_p=1.0,
)
print(response.choices[0].message.content)

For Anthropic-format clients, point the base URL at https://api.deepseek.com/anthropic and the rest of an existing Claude-style integration carries over largely unchanged.

Current official pricing, per million tokens:

Both models run a 1M-token context window with a 384K-token max output, and support thinking and non-thinking modes.

One date worth flagging if you've got an existing integration: the old deepseek-chat and deepseek-reasoner model names are being retired on July 24, 2026. Until then they're quietly routed to deepseek-v4-flash's non-thinking and thinking modes — but if your code still calls them by the old names, this is the month to update it.

Option 2 — Hugging Face (Official Open Weights)

If you want to run V4 yourself, there's exactly one place to get the real weights: the verified deepseek-ai organization on Hugging Face. Bookmark that link instead of searching for it every time — that single habit would have protected anyone from the fake GitHub repo above.

DeepSeek ships four V4 checkpoints:

The "-Base" checkpoints are pretrained only, without instruction tuning — useful if you're fine-tuning your own model, not what you want if you just want something to chat with. For everyday use, grab plain DeepSeek-V4-Flash or DeepSeek-V4-Pro.

To pull the weights with resumable, parallelized downloads:

pip install huggingface_hub hf_transfer
export HF_HUB_ENABLE_HF_TRANSFER=1
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash \
  --local-dir ./deepseek-v4-flash

Or the Python equivalent, which handles large multi-file repos more gracefully:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="deepseek-ai/DeepSeek-V4-Flash",  # swap for -Pro, -Flash-Base, or -Pro-Base
    local_dir="DeepSeek-V4-Flash",
    allow_patterns=["*.safetensors", "*.json", "*.py"],
)

FP8 weights run roughly one byte per parameter, so even Flash-Base is a 250–300GB pull before you touch anything else. Check the live "Files and versions" tab on the actual repo for the exact current size rather than trusting any fixed number printed in an article — this one included.

If Hugging Face is slow or rate-limited from where you're connecting, DeepSeek lists ModelScope directly on the official model card as an alternate mirror. That's the legitimate option — not a third-party mirror someone linked in a comments section.

The repository and weights are released under the MIT License. One thing worth knowing before you write inference code: V4 doesn't use a Jinja chat template the way V3 did. DeepSeek instead ships a Python encoding_dsv4 module in the repo's encoding/ folder that converts message lists into model input and parses the output back. Check that folder before hand-rolling a prompt format from memory of how V3 worked — it's a different system now.

For local deployment, DeepSeek's own recommended sampling settings are temperature = 1.0, top_p = 1.0. If you're using Think Max mode specifically, they recommend a context window of at least 384K tokens so it actually has room to use the reasoning budget.

Option 3 — Run a Quantized Build Locally (Ollama, llama.cpp, LM Studio)

The full FP8/FP4 release is built for multi-GPU hardware, not a single desktop. If you want something that fits on one consumer GPU, you're looking for a quantized build — and the safest way to find one is through the official model card, not a search engine.

On the DeepSeek-V4-Pro model page, there's a "Browse Quantizations" link under the deployment options, and the same exists on the Flash page. That surfaces community and vendor builds that explicitly declare deepseek-ai/DeepSeek-V4 as their base model — including versions packaged for llama.cpp, Ollama, LM Studio, and Jan. Going in through that link means you're seeing builds Hugging Face has actually linked back to the real repo, instead of whatever happens to rank first when you search "deepseek v4 ollama download."

That second part matters more than it sounds. Microsoft's research on the fake GitHub repo found it briefly outranking the real Hugging Face page for download-intent searches. The same SEO-poisoning pattern works just as well against "deepseek v4 ollama" or "deepseek v4 gguf" queries. Treat any download link you find through a search engine, rather than through huggingface.co/deepseek-ai directly, as unverified until you've confirmed who actually published it.

On the hardware side, vendors do release their own official quantizations for specific platforms — NVIDIA, for one, has published an NVFP4-quantized build of Pro tuned for its newest Blackwell-class GPUs. It's a decent example of what a legitimate vendor quantization looks like: posted under the vendor's own verified account, with serving instructions and benchmark data attached, not a bare .exe with a download button.

Hardware Requirements

Honest guidance, not "technically boots" numbers.

The reason Pro and Flash don't need proportionally more memory than V3-era models, despite the jump to a 1M-token context window, comes down to the new hybrid attention design DeepSeek calls CSA plus HCA. According to DeepSeek's own technical report, at 1M tokens of context, V4-Pro cuts KV-cache memory to roughly a tenth of what V3.2 needed at the same length, and Flash goes even further.

That said, this table changes every time someone publishes a better quantization. Check the live numbers on the deepseek-ai Hugging Face page before committing a weekend to a download.

Pick the access method that actually matches your hardware and your patience for a few-hundred-gigabyte download — the API is genuinely the fastest way to find out if V4 is worth that commitment before you make it. If you end up testing more than one access path, or comparing V4 against whatever you were running before, Macaron can hold onto your prompts, configs, and notes across sessions so you're not reconstructing your setup from scratch every time you switch. Try it free at macaron.im.

FAQ

Is there an official DeepSeek V4 repository on GitHub?

No. DeepSeek's real GitHub organization (github.com/deepseek-ai) hosts code and tooling, not the V4 model weights — those live on Hugging Face and ModelScope only. The "DeepSeek-V4" GitHub organization that briefly circulated download links was a fake created by attackers, documented by Microsoft Threat Intelligence in June 2026. Treat any V4 weights offered through GitHub as fake by default.

When did DeepSeek V4 actually release, and what's the difference between Pro and Flash?

DeepSeek released V4 as an open-weight preview on April 24, 2026. DeepSeek-V4-Pro is the larger model at 1.6T total parameters (49B activated per token); DeepSeek-V4-Flash is the faster, cheaper option at 284B total parameters (13B activated). Both support the same 1M-token context window.

What's the difference between the "-Base" models and the regular Pro/Flash releases?

The "-Base" checkpoints are the pretrained models before instruction tuning, meant for researchers and fine-tuning workflows, not direct chat use. If you just want to talk to the model or build an app on it, download DeepSeek-V4-Flash or DeepSeek-V4-Pro, not the -Base versions.

Is DeepSeek V4 actually open source — what's the license?

Yes. The repository and the model weights are both released under the MIT License, confirmed on the official Hugging Face model card.

Does V4 use the same chat template as V3?

No, and this trips people up if they're porting old V3 code. V4 doesn't ship a Jinja chat template at all. DeepSeek provides a Python encoding_dsv4 module instead, in the repo's encoding/ folder, for converting messages to model input and parsing output. Check that folder rather than reusing a V3-style prompt format.

Can I run DeepSeek V4 on a single consumer GPU?

For Flash, realistically yes, once a quantized build is available through the model card's "Browse Quantizations" link — its 13B activated parameters per token keep inference manageable even though the total weight footprint is large. For Pro, it's a much heavier lift: 49B activated parameters and a 1.6T total footprint mean aggressive quantization plus CPU offload at best on a single card, with noticeably slower responses than on multi-GPU hardware.

My Hugging Face download keeps failing or timing out — what's the fix?

Enable hf_transfer for faster, more stable large-file downloads (pip install hf_transfer, then export HF_HUB_ENABLE_HF_TRANSFER=1), and use snapshot_download rather than pulling files one at a time — it resumes properly on interruption. If Hugging Face is consistently slow from your location, use the official ModelScope mirror linked on the model card instead of searching for a third-party alternative.

From next article:

DeepSeek V4 Version History: V3 → V3-0324 → V4 Timeline (2026)

DeepSeek V4 Context Window: 128K vs 1M Tokens

DeepSeek V4 API: Rate Limits, Auth & Quickstart (2026)

DeepSeek V4 Architecture: MoE & Latent Attention Explained

How to Build an AI Agent with DeepSeek V4