DeepSeek and Open-Weight Models for Small Businesses: Where to Start

A closed model like ChatGPT or Claude is only available through its company’s app or API. An open-weight model’s parameters are released publicly, usually under a licence that allows commercial use with some conditions. You can run it on your own computer, on a rented cloud server or through a third-party host.

Open-weight is not quite the same as open source. The training data and full training code are often not published, and licences differ. Llama’s licence, for example, has its own terms, while some Mistral and Qwen models use Apache 2.0. Read the licence for the exact model you pick.

DeepSeek drew wide attention because its models performed close to leading closed systems on many benchmarks while being released openly. Its own hosted app runs on servers in China, which matters for data decisions, but the weights themselves can be run anywhere.

Most small businesses do not need to run their own model. A subscription to a mainstream assistant covers emails, social posts and quotes. Open-weight models become interesting in three situations.

• Privacy: a clinic, law office or accountant handling client documents may prefer that text never leaves its own machine.
• Volume: a shop generating thousands of product descriptions or translating a catalogue can cut per-request costs sharply.
• Customisation: a business with its own terminology, such as a spare-parts supplier, can fine-tune a small model on its catalogue.

If none of these apply, a closed model through a normal subscription is simpler and often better at writing.

There are three entry points, ordered from easiest to hardest.

Use a hosted provider

Services such as Groq, Together AI, Fireworks and OpenRouter run open-weight models and charge per use through an API. You get the lower cost without buying hardware. Check where the provider’s servers are and what it does with your data.

Run a small model on a laptop

Tools like Ollama and LM Studio let you download a model and chat with it locally in a few clicks. A recent laptop with 16 GB or more of memory can run smaller models comfortably. They are slower and less capable than the largest models, but good enough for summarising documents or drafting standard replies.

Rent a cloud server

For larger models or business tools used by several staff, you can rent a GPU server and run the model with software such as vLLM. This needs someone technical, either on staff or as a contractor.

Costs depend heavily on the route. Running a small model on an existing laptop costs nothing beyond electricity. Hosted API pricing for open-weight models is generally lower than closed frontier models for similar tasks, though prices change often, so check current rates.

Renting GPU servers is where costs climb. Hourly rates for capable GPUs add up quickly if the machine runs all month, so this route only makes sense for steady, high-volume work. A useful test: if your monthly bill on a hosted API is still small, self-hosting will not save you money once you count the time spent maintaining it.

Open-weight models carry the same weaknesses as any language model. They can state wrong facts with confidence, and smaller versions make more mistakes in languages other than English. Always check output before it reaches a customer.

Some models have been trained with content filters reflecting their country of origin, which can affect answers on political or historical topics. For product descriptions or appointment reminders this rarely matters, but it is worth knowing.

Security is your responsibility when you self-host. A model server exposed to the internet without a password is an open door. Keep it on a private network or behind authentication.