DeepSeek open weights
If you searched Hugging Face or deepseek-ai, you want checkpoints — not a chat box. Weights live on the Hugging Face org deepseek-ai. This site does not host them.
This site does not host weights. Model cards, licenses, and inference notes follow Hugging Face.
Weights vs the hosted API
The Hugging Face slug deepseek-ai is why many developers search “DeepSeek AI”. That org publishes open weights such as DeepSeek V4 Pro / Flash. The DeepSeek API is a hosted endpoint. You do not need those files to call it. Claude Code, Harness, and Deep Code default to the API, not to local weights.
Fetch weights only if you self-host, fine-tune, or must keep inference air-gapped. Licenses follow the model card — MIT where the card says MIT. Otherwise the web app or the API is less work than storing checkpoints.
When weights are the right door
Stay on Hugging Face when:
- You run vLLM, SGLang, Ollama or similar on your own GPUs.
- Compliance says data cannot leave your network — air-gapped inference.
- You need the published checkpoint for research or a fine-tune, under the license on that card.
- You already know the API is the wrong door because you must own the runtime.
What this page is not
Do not mix the org name with a hosted endpoint:
- Not a weight host. This site does not store or mirror checkpoints.
- Not the DeepSeek API. Calling https://api.deepseek.com/v1 does not require a download.
- Not a substitute for the model card. License and inference notes on Hugging Face win.
- Not the default path for Claude Code, Harness, or Deep Code — those tools call the API.
Source snapshot
- Where files live
- huggingface.co/deepseek-ai
- This site
- Does not host checkpoints
- Hosted API
- No weights required to call it
- License
- Follow the model card (MIT where published)
- Fetch when
- Self-host, fine-tune, or air-gapped
- Agent tools
- API by default, not local files
Where to go
Do not download from random mirrors:
- Open huggingface.co/deepseek-ai and pick the V4 model card you need.
- Read the license and inference notes on that card before you pull files. MIT applies only where the card publishes MIT.
- If you actually wanted a hosted call, stop and use the DeepSeek API. If you only wanted a laptop chat, use the web app.
- After the files are local, follow this site’s local-deploy guide for Flash-class serving — and that runtime’s README for flags.
Weights vs other doors
One brand, different jobs. Open the layer you actually need:
| Door | Role | When to open this page |
|---|---|---|
| Hugging Face deepseek-ai | Official-org checkpoints and cards | You need the files themselves |
| DeepSeek API | Hosted endpoint, no weights to store | You are calling a model, not serving one |
| Web app | Browser trial, no GPU | You only wanted to chat |
| Local deploy | Your GPU after the download | Files are already on disk |
| Claude Code / Harness / Deep Code | Agent shells on the API | You do not need local weights |
Boundaries and traps
- deepseek-ai is an org name on Hugging Face, not a reason to download before you call the API.
- This site is not a mirror. If a third-party zip disagrees with the card, the card wins.
- Claude Code, Harness, and Deep Code default to the API. Pointing them at localhost is a later, optional step after you serve weights yourself.
- License is per card. Do not assume every DeepSeek repo is MIT; read the card you are cloning.
Related guides
FAQ
Are weights the same as DeepSeek AI?
No. DeepSeek AI is the company / product name people search. deepseek-ai is the Hugging Face org slug that publishes checkpoints. Weights are the files on those model cards. The hosted DeepSeek API is a different door: you call an endpoint; you do not store the same files.
Do I need weights to call the DeepSeek API?
No. The API is a hosted endpoint. You send a model ID such as deepseek-v4-pro or deepseek-v4-flash and a key from the official open platform. Downloading huggingface.co/deepseek-ai is only for self-host, fine-tune, or air-gapped inference.
Do I need weights to use Claude Code, Harness, or Deep Code?
No. Those tools default to the DeepSeek API, not to local checkpoints. Get a key, point the tool at the hosted endpoint, and leave Hugging Face closed unless you later decide to serve weights yourself and retarget the tool.
Does this site host or mirror the checkpoints?
No. This site does not host weights. Cards, licenses, and inference notes on Hugging Face are the source of truth. Download from huggingface.co/deepseek-ai. We only explain which door to use.
What license are the weights under?
Follow the model card for the repo you clone. Where the card publishes MIT, treat it as MIT. Do not copy a license from a different DeepSeek repo or from this community page. If the card and a blog post disagree, the card wins.
When should I actually download weights?
Only when you must own the runtime: self-host on your GPUs, fine-tune, or keep inference air-gapped. If you wanted a conversation, use the web app. If you wanted a product integration, use the API. Fetching a checkpoint “just in case” is usually the wrong first step.
Where should I go if a command or flag breaks?
The Hugging Face model card and the runtime README (vLLM, SGLang, Ollama, or whatever you launched). This page is a map to the org, not the issuer of weights, licenses, or serve flags. After the files are local, use the local-deploy guide on this site for the Flash-class path.