Introduction to Open Weight Models: What You Actually Download (1/5)

About this series

Open weight models have gone from research curiosity to a line item in nearly every enterprise AI strategy I review. Yet most of the conversations I have with CIOs and architects still collapse into a false binary — “open source vs. OpenAI” — that misses what actually matters: control, cost, risk and fit-for-purpose.

This five-part series is my attempt to fix that framing. Each post stands on its own, but together they build from first principles to a working enterprise blueprint:

1. Introduction to Open Weight Models (1/5) — what you actually get when you download a model, and why “open weight” is not the same as “open source.”
2. Open Weight vs. Closed Models (2/5) — an honest comparison across capability, cost, control, security and operations.
3. Where Enterprises Use Each — and a Reference Architecture (3/5) — a workload-placement framework and the platform that makes a multi-model strategy work.
4. Open Weights in the Bank (4/5) — a worked example of a deployment at a large retail and commercial bank.
5. The Future of Open Weights (5/5) — where this is heading and what leaders should do in the next 18 months.

The audience is technical leaders — architects, platform owners, CIOs and CTOs. I’ll go deep enough to be useful, but this is not a tutorial on transformer math as I am not a data scientist 🙂

Part 1

This is Part 1 of a five-part series on open weight models in the enterprise. Before we can argue about open vs. closed, or design a reference architecture, we need to be precise about what an open weight model actually is — because the term gets used loosely, and loose terms lead to bad procurement and worse risk decisions.

A model is simply a very large set of numbers

A large language model is, stripped of hype and geopolitical noise, a neural network: layers of simple units (“neurons”), each taking a weighted sum of its inputs, adding a bias, and passing the result through an activation function. Each connection between two neurons has a weight. There is a bias for each neuron. These are the parameters of the model.

Training is simply the (very expensive) process of finding good values for those parameters. When you hear “a 70B model”, it means 70 billion learned numbers. Stored with 16 bit precision, that’s about 140GB of data – and that file, with a little bit of configuration, is the model.

Modern LLMs are transformers, rather than the simple feed-forward network in the diagram below, but the principle is the same. Attention heads and feed-forward blocks are all parameterized by weight-matrices and bias-vectors. The model is owned by whoever has those matrices.

What a model is, and what an open-weight release actually contains

So what does “open weight” mean?

An open weight model is one where the model publisher releases the trained parameters for anyone to download, along with the architecture config and tokenizer needed to run them, under a license that says what you may do with them.

In contrast, with a closed model, you rent access to a capability over an API, and the provider decides which version you get, where it runs and when it is retired. With an open weight model, you hold the asset. You can:

• Run it anywhere—on-premises, in your own VPC, at the edge, or in an air-gapped enclave in your sovereign cloud.
• Adapt it—fine-tune it on your data, distill it into something smaller, or quantize it to 8- or 4-bit so it fits cheaper hardware.
• Pin it—freeze an exact version for as long as your validation and audit processes require.
• Inspect it—probe activations, and benchmark it on your workloads rather than the vendor’s.

Open weight is not open source

This is the distinction one most sees often see blurred at the business leadership level. Releasing weights is not the same as releasing the recipe. Most open weight releases do not include the training data, the full training code, or the preference data used for alignment. You can use and modify the model; you generally cannot reproduce it.

The Open Source Initiative made this explicit with its Open Source AI Definition, which requires detailed information about the training data and the code used to train and run the system — not just the weights. By that bar, only a handful of research-grade releases are truly open source. Everything else sits somewhere on a spectrum:

The license column is the one your legal team should care about most. A “community” license with an acceptable-use policy or a user-count threshold is a very different contractual object from Apache 2.0.

Where things stand in late 2026

Two years ago, open weight models were roughly a generation behind the frontier. That gap has narrowed sharply. DeepSeek V4-Pro, released in April 2026, is a 1.6-trillion-parameter Mixture-of-Experts model that activates only 49 billion parameters per token, ships with a 1M-token context window, and matches or beats closed frontier models on several coding benchmarks. DeepSeek itself describes its position as roughly three to six months behind the absolute frontier (Digital Applied).

The gap has not closed uniformly, though. The best open models still trail by roughly 5 to 18 points on general knowledge, graduate-level science and long-context retrieval benchmarks. And the field moves monthly — Qwen, Mistral, Moonshot and Z.ai have all shipped significant checkpoints since the summer (Kingy AI).

Three ideas to carry into the rest of this series

1. Weights are an asset you hold, not a service you rent. That shifts both control and responsibility to you.
2. The license is the real contract. Read it before the benchmark chart.
3. Parameter count is not cost. With Mixture-of-Experts architectures and quantization, active parameters and memory footprint drive your infrastructure bill — not the headline number.

In Part 2, I’ll put open weight and closed models side by side and look at where each genuinely wins — and where the conventional wisdom is wrong.

Featured image designed by Magnific