DeepSeek: frontier-sized open weights under the MIT licence

Contents

1. What it does for a company
2. Strengths and limits
3. Pricing
4. Questions to ask before choosing DeepSeek
5. Sources

Type	Open-weight model family
Pricing	Published
Website	huggingface.co

DeepSeek releases both the code and the weights of its V4 generation under the MIT licence, which places no user caps, revenue thresholds or naming rules on a company that uses them⁠12deepseek-ai/DeepSeek-V4-Pro-0813DeepSeek (Hugging Face)deepseek-ai/DeepSeek-V4.1-FlashDeepSeek (Hugging Face).

Best for: Coding and agent work under the most permissive licence among the frontier-sized families.

Not the right fit for: Companies that want a small model or a European developer.

What it does for a company

Two models make up the current generation. DeepSeek-V4-Pro is a mixture-of-experts model with 1.6 trillion parameters, 49 billion of them active per token, pretrained on more than 32 trillion tokens with a one-million-token context⁠3deepseek-ai/DeepSeek-V4-ProDeepSeek (Hugging Face). DeepSeek-V4.1-Flash is lighter in use: a 552-billion-parameter backbone that activates about 8 billion parameters while reading a prompt and 16 billion while writing, also with a one-million-token context and with image input⁠2deepseek-ai/DeepSeek-V4.1-FlashDeepSeek (Hugging Face).

For coding and agent work the vendor-reported results are the strongest found in this review: 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1 for V4.1-Flash⁠2deepseek-ai/DeepSeek-V4.1-FlashDeepSeek (Hugging Face). These are DeepSeek’s own figures and should be checked against a company’s own tasks.

The cost is hardware. The published vLLM serving example for V4-Pro uses a single node of four Nvidia GB300 GPUs⁠1deepseek-ai/DeepSeek-V4-Pro-0813DeepSeek (Hugging Face). Per-token endpoints avoid that: Together AI lists V4.1 Flash at $0.30 per million input tokens and $1.20 per million output tokens⁠4Pricing: serverless inference, dedicated inference and GPU clustersTogether AI. The model cards do not include a fine-tuning guide.

Strengths and limits

Strengths

• MIT licence with no user caps, revenue thresholds or naming rules.
• Vendor-reported results lead this group on agentic coding: Terminal-Bench 2.1 at 90.6 for V4.1-Flash.
• One-million-token context on both current models.

Limits

• Large: the published vLLM example uses a node of four GB300 GPUs.
• We found no official fine-tuning guide on the model cards.
• Policies that exclude Chinese developers apply here too.

Choose another option if: You need to fit on one GPU (consider gpt-oss or Gemma 4) or want a European vendor (Mistral).

Pricing

Weights are free. Together AI lists V4.1 Flash at $0.30 per million input tokens and $1.20 per million output tokens.

Questions to ask before choosing DeepSeek

• Does our security policy permit weights from a Chinese developer when they run on hardware we control?
• Do our own coding and agent tasks show the same lead that the vendor-reported benchmarks do?
• Will we serve V4-Pro ourselves, or is V4.1-Flash through an endpoint enough for our volume?
• Which fine-tuning tools will we use, given that we found no official guide?

See how DeepSeek compares with the other options in the main guide.

Sources

1. deepseek-ai/DeepSeek-V4-Pro-0813. DeepSeek (Hugging Face). Accessed 9 Nov 2026.
2. deepseek-ai/DeepSeek-V4.1-Flash. DeepSeek (Hugging Face). Accessed 9 Nov 2026.
3. deepseek-ai/DeepSeek-V4-Pro. DeepSeek (Hugging Face). Accessed 9 Nov 2026.
4. Pricing: serverless inference, dedicated inference and GPU clusters. Together AI. Accessed 9 Nov 2026.

Further reading

• Buyer's guide · Updated quarterlyThe best open-source AI in 2027, for companies that want to own itAn analysis of the open-weight model families a company can run today and the layer that turns one into its own, compared on seven published criteria, with sourced licences, costs and the established path from rented AI to owned AI.
• Licence guideOpen source vs open weights: what each licence lets a company doAlmost every model sold as open-source AI publishes its weights but not its training data. The difference rarely matters for internal use. It matters a great deal for companies that resell a model or build a product on it.
• Case analysisGoing headless like Revolut: what replacing Claude Code with an open model tookRevolut now runs most of its AI coding on an open-weight model it serves itself, for about a tenth of its former bill. A reconstruction of what it built, what the numbers mean for other companies, and the step that comes after.

Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski