MMOOZZGG.com

How to build an AI agent

How to build your own neural network: an agent on an open model, not training

5 min read7 sources

A company that asks how to build its own neural network usually needs an AI that knows its documents and runs on its own server. For that, you do not train a model from scratch: full training of the open model DeepSeek-V3 took 2.788 million GPU hours on H800 chips. You build your own neural network as an agent on a ready open model.

On this page · 4 sections

Build it yourself in a week

Training a model versus building an agent

To train a model means to run trillions of words through it so it learns language and general knowledge. DeepSeek-V3 was trained on 14.8 trillion tokens (word fragments); the model has 671 billion parameters, of which 37 billion work on each token.

Parameters are the numbers inside a model that are set during training, and tokens are the pieces of words the model reads and writes text with. The more parameters, the more GPUs and time training takes.

A company does not need this work. Open models already have language and general knowledge, and the model gets knowledge about your company in another way.

Training from scratch makes sense for companies that sell the models themselves. They need their own data sets, a GPU cluster and a team of researchers, while a company that needs an assistant for its staff has a different task.

An agent is a ready model plus an instruction, your documents and tools. Documents are connected through RAG: before answering, the agent searches your knowledge base for the right fragments and answers from them. Lewis and co-authors described this approach in 2020, and it works without training the model.

The agent’s knowledge updates at once: replace a policy in the knowledge base, and the agent answers by the new one. A model trained on the old policy would have to be trained again.

So the answer to the question of your own neural network for business is almost always the same: an open model, your documents and an agent on top of them.

Open models with a free license

Open weights mean that you can download the model file and run it yourself. Two conditions matter for business: the license allows commercial use, and the model fits your hardware.

DeepSeek R1 came out on 20 January 2025 under the MIT license. The Qwen 3 family came out on 29 April 2025 under the Apache 2.0 license. Both licenses let a company use the models in its work without paying the developer.

Such models run in Ollama, a program under the MIT license: it downloads a model with one command and runs on your computer or server.

In Ollama you can also make “your own” model without training. A Modelfile sets the base model (the FROM line), a system message with the role and rules (SYSTEM) and the parameters (PARAMETER). The command ollama create name -f ./Modelfile builds a model under your name from them, and ollama run name starts it.

An example of such a file in two lines: FROM qwen3, and SYSTEM with the text “You are a sales consultant for our company; answer briefly and only from the price list.” After ollama create, the model answers in this role without reminders.

This is the same open model with your instruction inside, and staff work with it as the company’s neural network. A list of other open models with sizes and licenses is in the review of open models.

When you need fine-tuning

Fine-tuning changes a ready model on your examples. It is used when instructions and documents are not enough: you need a special answer format, narrow terminology or a style the model does not keep from an instruction.

An example: the model must sort incoming requests into a strict format for your accounting system, and an instruction with samples does not help. Then labeled requests and fine-tuning solve the task.

Full fine-tuning of a large model is expensive, so teams more often use LoRA: it trains a small add-on to the model and leaves the model itself untouched. The authors of the method showed that for GPT-3 175B it cuts the number of trainable parameters by 10,000 times and GPU memory needs by 3 times.

The order is this: first the instruction, then RAG with documents, and only if both fall short, fine-tuning. Fine-tuning needs good examples of correct answers and a person who checks the result.

Fine-tuning is tied to a model version. When a new version comes out, the work must be repeated, while the instruction and the documents move to the new model unchanged.

Your own neural network on your own server

Your own neural network in the full sense runs on your server: the model, the documents and the request log never leave the company. For this you need a server with a GPU or a computer with enough memory, Ollama and an agent on top of it.

How much memory you need depends on the model size: the more parameters, the more memory. Calculate the price of your own server for the model you need in the local AI calculator.

Start with a test on a laptop: install Ollama, take a small model and ask it your staff’s typical questions. If the answers work for you, move to a server; if not, test a larger model before you buy hardware.

You can build such a neural network in a week without a programmer: Ollama, the model and a no-code agent builder install by their documentation. The step-by-step guide is on the local agent page.

You need a programmer when the agent must work with the CRM, the ERP and clients’ personal data. Then access rights, an action log and human review of answers are added to the model.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Alpine Linux packages, 2026 (pkgs.alpinelinux.org)
  2. Simon Willison, 20 January 2025 (simonwillison.net)
  3. Simon Willison, 29 April 2025 (simonwillison.net)
  4. arXiv 2005.11401, 22 May 2020 (arxiv.org)
  5. arXiv 2412.19437, 27 December 2024 (arxiv.org)
  6. arXiv 2106.09685, 17 June 2021 (arxiv.org)
  7. Ollama Docs, 2026 (docs.ollama.com)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us