Open weight models are AI models whose trained parameters are available to download and run on infrastructure you control. Depending on the license and model, your team can inspect, fine-tune, and deploy them without sending every request through a commercial API. That can improve control and portability, but it does not automatically reduce cost.
Open weights are becoming a boardroom topic because they change where companies pay for AI and what they can own. Instead of paying only through a model provider's API, a company can take responsibility for compute, deployment, security, evaluation, and maintenance. This guide explains what open weights provide, what they do not provide, and how to decide whether they fit your use case.
Weights are the learned numerical parameters inside a neural network. Together, they shape how the model recognizes patterns and generates an answer. They are not simple controls such as a single dial for location, recency, or relevance.
A pizza search still offers a useful analogy. A useful answer should account for location, cuisine, reviews, availability, and the user's intent. In production, teams usually influence those factors through prompts, retrieval, context, tools, routing, and fine-tuning rather than manually turning one isolated weight up or down.
An open weight model makes the trained parameters available under a license. Your team can host the model, test it, fine-tune it, and control when a new version enters production. A commercial API keeps the underlying parameters with the provider. You can shape outputs through the API, but you cannot independently host or version the model.
We have done this kind of weight work before. With DataFlik, we built a machine learning model that predicted 74% of all houses sold in the United States every month. The dataset had 1,400 columns. The job was understanding which weights were on or off under given conditions, and updating those weights monthly as the market shifted. That project is a proof point. Weight tuning is work we have shipped, not theory.
The model developer creates the initial weights during training. An organization with access to open weights can then fine-tune the model, train adapters, quantize it for different hardware, or keep the original model unchanged and shape its behavior through the surrounding application.
With a commercial API, the provider controls the model and its release schedule. Your team controls prompts, tools, retrieval, and application logic, but the provider can still change the model behind the endpoint.
With open weights, your team controls which version runs in production and when it changes. That makes rollback, testing, and deployment more predictable. It also transfers responsibility for infrastructure, evaluation, security, and maintenance to your organization or implementation partner.
A harness is the software around a model. It can manage prompts, memory, retrieval, tools, permissions, routing, evaluations, and user context. It usually does not change the model's weights during each request. It changes the information and instructions the model receives, then decides what should happen with the response.
This distinction matters because the model is only one part of a production system. Two products can use the same underlying model and produce different results because their retrieval, context management, tools, and evaluation loops are different.
Harnesses can also be specific to a workflow or industry. A legal system may retrieve approved case law and restrict which sources the model can cite. A real estate system may combine location, property, and market data while enforcing role-based access. Open-source components can reduce build time, but they still need to be evaluated, secured, and adapted to the company's data and operating rules.
The AI stack has four layers: chips, cloud infrastructure, the model and API layer, and the application layer. Margins are high at every level, though the exact numbers vary by source and reporting period.
| Stack layer | Example companies | Approximate gross margin |
|---|---|---|
| Chips (GPUs) | NVIDIA | ~75% |
| Cloud infrastructure | AWS, Google Cloud, Azure | High (not publicly broken out for AI) |
| Model / API layer (per-token) | OpenAI, Anthropic | Not yet proven as a profit center |
| Application layer (SaaS) | HubSpot, Notion, Cursor | 80 to 90% |
NVIDIA's GAAP gross margin for fiscal 2026 was 71.1%, with quarterly margins reaching 75%. Application-layer SaaS companies like HubSpot and Notion typically run at 80 to 90% gross margin. Cursor built a fast-growing business by wrapping Claude and OpenAI models in a better code-editing experience, collecting SaaS margins on top of someone else's model. Cloud providers do not break out their AI-specific margins publicly, but the numbers are high.
The layer that is new and unproven is the per-token payment layer between infrastructure and the application. Companies like OpenAI and Anthropic charge per token, and their margins are not yet established as durable profit centers. The economics are still settling.
If a company owns its compute, whether cloud or bare metal, a platform provider can compress its own token margin to win customers. Say a provider drops from 90% margin to 50% on tokens to undercut a competitor's harness. The customer gets cheaper inference. The provider still makes money on the infrastructure. The competitor's harness, built on the same model, loses its pricing power.
Why pay HubSpot or Notion to run OpenAI prompts when you could run the same prompts on your own dedicated instance? The SaaS company is collecting 80 to 90% margin on top of someone else's model. If you own the model and the harness, that margin is yours.
Distillation is the process of training a smaller model to replicate the behavior of a larger one. In software terms, it is reverse engineering. You study what the large model produces, train a smaller model to produce similar outputs, and release it.
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first open model, Inkling, in July 2026. The company acknowledged using other open-weight models, including Chinese model Kimi K2.5 from Moonshot AI, to help generate some of its early post-training data. They pre-trained the model from scratch, but the post-training process drew on existing open weights. This is close to what the industry calls distillation, and it means the barrier to building a competitive model is lower than most people assume.
Open source models are being released openly through distillation. The models that result are not as capable as the frontier models from OpenAI or Anthropic in every benchmark, but they are close enough for many production use cases. And they are free to download and run.
There is a cost trade-off. Open source and local models can burn more tokens than commercial APIs for the same task, depending on how they are configured and prompted. The latency and total cost can land in the same place even with no per-token bill. You are paying in electricity and compute time instead of paying per token. The savings depend on your volume and your engineering capacity to optimize the model.
Open weights give a company another architecture option. They can reduce dependency on one model provider, keep sensitive workloads inside a controlled environment, and choose when a model version changes. They can also create a larger operational burden than an API.
The strongest ownership strategy is model agnostic. Keep your data, retrieval, evaluations, permissions, workflows, and application logic under your control so the underlying model can change when the economics or performance justify it. A model swap is never automatic, however. Each candidate still needs testing for accuracy, latency, safety, and compatibility with existing workflows.
This is where Owned Intelligence applies. The goal is to build an intelligence layer that reflects the company's data, rules, and successful decisions while fitting its existing stack. Open weights can support that goal, but they are one architectural choice rather than the goal itself.
Open weights deserve serious evaluation when several of these conditions are true:
A commercial API is often the better choice when usage is low or unpredictable, speed matters more than control, the frontier model is required, or the company does not have a team responsible for production operations. Paying per token can be cheaper than carrying infrastructure and maintenance before utilization reaches a meaningful level.
Use this checklist with IT, Legal, Finance, and the business owner before choosing an architecture:
Run the economics before choosing the architecture. Use the LLM Cost Calculator to compare API and self-hosted costs at your expected volume. Then use the Self-Hosted vs. Third-Party LLMs guide to account for engineering time and operational ownership.
If the numbers support ownership but your team lacks the capacity to implement it, talk with NineTwoThree about validating one scoped use case before committing to a broader architecture.
Open weights means a model's trained parameters are available to download under a license. An organization can run the model on infrastructure it controls and, when the license allows it, fine-tune or modify it for a specific use case.
The model developer creates the initial weights during training. A team with access to open weights can later fine-tune the model or train adapters. Many production teams leave the base weights unchanged and improve results through retrieval, prompts, tools, and application logic.
Usually, no. A harness manages prompts, memory, retrieval, tools, routing, permissions, and evaluations around the model. Fine-tuning changes model parameters through a separate training process.
They may be free to download, but running them is not free. The company still pays for compute, storage, engineering, monitoring, security, and maintenance. The right comparison is total cost at the expected workload, not API price versus a zero-dollar model license.
They can provide greater control over where data is processed and stored. Security still depends on the deployment, access controls, patching, audit logs, monitoring, and operating practices around the model.
Yes, if the surrounding system was designed for portability. Every replacement model still needs to pass the same quality, safety, latency, and cost evaluations before it enters production.
Find out whether open weights fit your AI use case. Start with the evaluation checklist above, then run your expected volume through the LLM Cost Calculator. If the economics support ownership and you need a team to validate the architecture, talk with NineTwoThree about one scoped use case.
.jpg)