Is DeepSeek Open Source? A Clear Look at Licensing in 2026

Short answer, up front: yes, DeepSeek open source releases are real, and the current generation — DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro-0813, the two models behind DeepSeek’s API today — ships both weights and code under the MIT License. But “open source” in AI is a term people use loosely, and older DeepSeek models (V3 base, Coder-V2, VL2) split a permissive code license from a more restrictive weights license. If you plan to fine-tune, self-host or ship a commercial product on top of DeepSeek, those differences matter.

This guide walks through exactly what’s open as of October 2026, what each license allows, where to download the weights, and what trade-offs still apply compared with closed APIs from OpenAI, Anthropic and Google.

Is DeepSeek open source? The short answer

Yes — with caveats worth knowing. DeepSeek publishes its models as open-weight releases on Hugging Face, and every recent release (V4.1-Flash-1017, V4-Pro-0813, V4.1-Flash, V3.2, V3.1, R1) carries the MIT License on both code and weights. DeepSeek the company runs a paid API and a free chat app, but the underlying models are downloadable, forkable and commercially usable.

That makes DeepSeek’s recent releases more permissive than Meta’s Llama 4, whose community license requires companies with more than 700 million monthly active users to request a separate license from Meta, and more open than OpenAI’s GPT-5, Anthropic’s Claude and Google’s Gemini models, whose weights are not published. It does not automatically make every DeepSeek artefact MIT-licensed — some older model weights sit under a separate custom license. The table below shows the split; always check the Hugging Face repo for the specific model you plan to ship.

DeepSeek licensing, per model

There are two licenses in play across the DeepSeek catalogue:

• MIT License — a short, permissive license that allows commercial use, modification, redistribution and sublicensing, as long as the copyright and permission notice is kept in all copies or substantial portions.
• DeepSeek Model License (DeepSeek License Agreement, version 1.0, 23 October 2023) — a custom license with use-based restrictions (Attachment A lists, for example, unlawful use, military use and harming minors) but otherwise broad freedoms; the model cards of V3, Coder-V2 and VL2 each state that the series supports commercial use.

Here is where each major model lands, as stated on its Hugging Face repo (checked 24 October 2026):

For the current models, the Hugging Face cards for DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro-0813 state plainly that the repository and model weights are licensed under the MIT License. The same is true for R1: DeepSeek’s January 2025 launch notes said code and models were released under the MIT License (“Distill & commercialize freely!”), and the R1 card adds that the R1 series supports commercial use and allows modifications and derivative works, including distillation for training other LLMs.

What MIT actually grants you

For MIT-licensed releases such as V4.1-Flash-1017, V4-Pro-0813 and R1, MIT means you can:

• Download the weights and run them on your own hardware.
• Fine-tune them on proprietary data and keep the result private.
• Serve the model to paying customers without a royalty or revenue share.
• Distill them into smaller models — the R1 card names distillation explicitly.
• Redistribute modified versions under other terms, including proprietary ones.

The one hard requirement is keeping the MIT copyright and permission notice in all copies or substantial portions. If you are auditing DeepSeek privacy and data handling for a contract, note that the license covers the weights you run yourself; the hosted chat and API are governed by DeepSeek’s Terms of Use and Privacy Policy instead.

What the DeepSeek Model License adds

For older weights (V3 base/chat, Coder-V2, VL2), the DeepSeek Model License still permits commercial deployment and derivative work, but layers in use-based restrictions. Its Attachment A rules out, among other things, use that violates applicable law, military use in any way, exploiting or harming minors, generating false information to harm others, defamation or harassment, fully automated decision-making that adversely affects someone’s legal rights, and discrimination based on legally protected characteristics. Section III requires anyone who distributes the model or a derivative to include the same use-based restrictions as an enforceable provision and to tell downstream users about them.

Where to download DeepSeek weights

DeepSeek publishes official weights on two platforms:

• Hugging Face — the primary distribution channel under deepseek-ai/, for example deepseek-ai/DeepSeek-V4.1-Flash-1017, deepseek-ai/DeepSeek-V4-Pro-0813 and the entire back-catalogue.
• ModelScope — the Chinese-hosted equivalent, useful if Hugging Face is slow from your region. DeepSeek’s 17 October 2026 release notes link both platforms for DeepSeek-V4.1-Flash-1017.

Community builds appear on Ollama and in quantized GGUF form, but as of October 2026 Ollama’s deepseek-v4.1-flash and deepseek-v4-pro tags are cloud-only. If you are looking for something you can actually run on a workstation rather than a multi-GPU server, start with the R1 distilled models (1.5B to 70B) or the Coder family. The current flagship repositories hold about 510 GB (V4.1-Flash-1017) and 893 GB (V4-Pro-0813) of files — data-centre workloads, not laptop workloads.

How V4 compares on openness

The current releases continue V4’s frontier-tier open-weight approach. DeepSeek does not state parameter counts for DeepSeek-V4.1-Flash-1017 or DeepSeek-V4-Pro-0813; the often-quoted figures (V4-Pro 1.6T total / 49B activated, V4-Flash 284B / 13B, both with a one-million-token context) come from the April 2026 preview card, which also reported that in the 1M-token setting the preview V4-Pro needed only 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2. The V4-Pro-0813 card says that model is built on the preview’s structure with an added DSpark speculative-decoding module.

On capability, DeepSeek’s own numbers are the main reference. The V4.1-Flash-1017 card (17 October 2026) reports, for example, 77.0 on DeepSWE v1.1 against 74.0 for Opus 5 and 73.0 for GPT-5.6 Sol in the same table, while Opus 5 leads on HLE without tools (56.3 against 42.6). These are DeepSeek-reported results — treat them as a starting point, not a final verdict.

How it stacks up against other “open” releases

License details checked on the vendors’ own pages and Hugging Face repos on 24 October 2026. For a deeper head-to-head, see DeepSeek vs Llama on licensing specifics. The practical read: with DeepSeek’s recent releases you can take the weights, fine-tune, and ship a commercial product without a signed agreement.

Is DeepSeek “fully” open source? The honest answer

By the Open Source Initiative’s Open Source AI Definition (version 1.0), a fully open AI system also makes available detailed information about its training data, the complete source code used to train and run it, and its parameters, all under OSI-approved terms. DeepSeek publishes technical reports, weights, inference code and MIT licensing, but it does not publish:

• The pre-training dataset, or a complete description of it, in a reproducible form.
• The full training pipeline source.
• Exact data-mixing recipes.

That puts DeepSeek in the “open-weight, not fully open” bucket by OSAID criteria. For most practical users — developers fine-tuning for a product, researchers studying capabilities, enterprises self-hosting for privacy — open weights plus MIT is what you actually need. If reproducibility is a hard requirement, check whether a release publishes its training data information and training code, not just its weights.

Accessing the models — open weights vs API

Even if you never download the weights, you can use the same models through DeepSeek’s managed API. As of October 2026 it serves two model IDs: deepseek-flash (DeepSeek-V4.1-Flash-1017, the only one with image input) and deepseek-v4-pro (DeepSeek-V4-Pro-0813). Both are also published as MIT-licensed open weights on Hugging Face.

Chat requests hit POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com; DeepSeek also offers an Anthropic-compatible interface at https://api.deepseek.com/anthropic. The API is stateless: unlike the web chat and mobile app, which keep your chat history, API clients must resend the full conversation on every request. Thinking mode is a request parameter on either model, on by default, and returns reasoning_content alongside the final content. See the DeepSeek API documentation for setup, code samples and the parameter reference.

API pricing snapshot

Because the weights are open, there is no vendor lock on inference — you can always run the models yourself or pick a third-party host. DeepSeek’s own rates as of October 2026, in USD per million tokens from its official pricing page; off-peak rates are half the peak rates, and peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays:

Verify current rates on the official pricing page before committing to a budget; DeepSeek says prices may vary. See our DeepSeek API pricing breakdown for worked examples, or the DeepSeek pricing calculator if you want to cost out your own workload.

Running DeepSeek open source weights locally

Open weights only matter if you can actually run them. The hardware reality is blunt: the current V4.1-Flash and V4-Pro releases are large models that assume multi-GPU inference. If you want something on a single machine, drop down the range — the sizes below are the default (4-bit quantised) Ollama downloads for each distilled model as of October 2026, and you need at least that much free memory plus headroom:

1. R1 Distill 1.5B – 14B — 1.1 GB to 9.0 GB downloads; the realistic choice for laptops.
2. R1 Distill 32B — a 20 GB download; a high-memory GPU or a well-equipped workstation.
3. R1 Distill 70B — a 43 GB download; serious workstation hardware.
4. Full R1, V3.2, V4.1-Flash, V4-Pro — data-centre only, or managed cloud inference.

Licensing note: the R1 card says the Qwen-based distills derive from Qwen2.5 (originally Apache 2.0) and the Llama-based ones from Llama 3.1 and Llama 3.3, which carry Meta’s licenses. The simplest on-ramp is Ollama. Follow our guide to running DeepSeek on Ollama for a one-command install of a distilled variant. If you prefer containerised deployment, our install DeepSeek locally walkthrough covers Docker and direct Python paths. Confirm your machine meets the minimums in the DeepSeek system requirements reference before downloading tens of gigabytes of weights.

Why does DeepSeek give the weights away?

It is a fair question, and the likely answer is a mix of strategy and engineering culture. Open-weight releases drive adoption, surface bugs through community use, build developer mindshare in a space dominated by closed labs, and — not incidentally — put pressure on closed competitors’ pricing. R1’s MIT release in January 2025 drew wide attention, and the V4 and V4.1 releases continue that playbook with frontier-tier models and published efficiency claims.

That doesn’t make DeepSeek a charity. The company monetises through its paid API, which is where much commercial demand lands. Open weights are a lever, not the business.