This article is part of our AI Agent Guides. See how models compare this week on the AI Agent Leaderboard, a ranking of rankings refreshed every week.
Choose a proprietary agent if you want the fastest start with the least upkeep. Consider open weights if data control, customization or high-volume cost outweighs the extra work. Most small and midsize businesses should start with a hosted service, then revisit this once they know their volume and their data rules.
The term also misleads people. "Open weight" does not mean free of terms, and it does not automatically mean your data stays private.
Disclosure: eSEOspace is a member of the Claude Partner Network.
Key Takeaways
- Open-weight means you can download the trained model's parameters. Proprietary means you use it only through the vendor's service or products.
- Open weight is not the same as open source. The Open Source Initiative treats them differently.
- Licenses vary widely, even inside a single vendor's lineup. Read the license for each exact model.
- Data control comes from where the model runs, not from the label. An open model on someone else's API is still someone else's service.
- Check how open-weight models compare today with the open-weight filter on the AI Agent Leaderboard.
What the terms mean
Weights are the numbers inside a trained model, and they determine how it responds. Publishing them lets you run the model on hardware of your choosing. A proprietary model keeps those numbers private, so you reach it through an API or an app.
The Open Source Initiative says open weights do not include training code, the training dataset or full data transparency. Its Open Source AI Definition also asks for detailed data information. So a vendor saying "open" may mean less than you assume.
The word "agent" adds one more layer. An agent is a model plus a harness, so open models can run in closed harnesses, and the reverse. This guide is about the model choice, which is usually the larger decision.
The main differences
| Factor | Proprietary, via vendor | Open-weight, hosted by a provider | Open-weight, self-hosted |
|---|---|---|---|
| Setup effort | Low | Low | High |
| Where your data goes | Vendor | Hosting provider | Your systems |
| Who maintains it | Vendor | Provider | You |
| Upfront hardware | None | None | Possibly significant |
| Cost pattern | Per token or seat | Per token | Fixed infrastructure plus staff |
| Control over version | Vendor can retire it | Provider can drop it | You keep the file |
| Fine-tuning | Limited to vendor options | Varies | Full, within the license |
| Legal terms | Vendor contract | Provider terms plus model license | Model license |
The middle column surprises people, and for many companies it is the practical option. Mistral's documentation, for example, lists third-party open-weight models from DeepSeek and Z.ai on its own platform (checked on 2 November 2026).
Factor 1: data control
This is the most common reason for choosing open weights, and also the most misunderstood. A 2024 US government report from NTIA named this benefit plainly: open models let users run them without sharing data with third parties. That benefit only applies when you actually host the model yourself.
If you use an open model through an API, your data goes to that provider. Read its terms just as you would read a proprietary vendor's. A license on the model file tells you nothing about what a hosting company does with your prompts.
Proprietary does not mean careless, either. For example, Anthropic's commercial terms say it may not train models on customer content from its services. Terms differ by vendor and plan, so check yours.
If you handle regulated data, get a qualified professional to review the setup. Self-hosting moves responsibility onto you, and it does not satisfy any rule by itself.
Factor 2: licenses
Open-weight licenses range from very permissive to heavily conditioned. Here are examples from pages we checked on 2 November 2026.
| Model family | License | Terms worth noting |
|---|---|---|
| gpt-oss | Apache 2.0 | Permissive, and described as suited to commercial deployment |
| Mistral Small 4, Ministral 3 | Apache 2.0 | Permissive |
| Mistral Large 4 | Mistral Ecosystem Open Weight License | Separate license, so read it |
| DeepSeek V4.1 Flash, GLM 5.3 | MIT, as listed by Mistral | Check the publisher's own page too |
| Llama 4 | Llama 4 Community License | Must display "Built with Llama". Services over 700 million monthly users need a separate license |
| Gemma | Gemma Terms of Use | Includes a prohibited use policy |
Notice the Mistral row: two models from one company can have different licenses. Older families have the same pattern. The original Qwen repository uses Apache 2.0 for code, but it says to check the license with each model for commercial use.
Have a lawyer read the actual license before you ship a product on it. This guide is not legal advice.
Factor 3: hardware and running costs
Open weights are free to download, but running them isn't free. Hardware needs vary a lot, even within one family.
OpenAI's gpt-oss model card says its larger model fits on a single 80GB GPU, such as an NVIDIA H100. Its smaller model runs within 16GB of memory. Tools such as Ollama's library and vLLM help with running models, but they do not supply the hardware.
We have not given a break-even figure, because we could not find a reliable source that fits most businesses. It depends on your volume, your utilization and your staff. Work it out using this list:
- Hardware or cloud GPU rental, including idle time
- Staff time for setup, updates, monitoring and security
- Fallback capacity when a machine fails
- Cost per finished task, not just per token
Also compare hosted prices directly. DeepSeek's pricing page, for example, lists per-token prices that differ between peak and off-peak hours (checked on 2 November 2026). Token price is not cost per task, though.
Arena reports that some models cost more in practice despite cheaper list prices.
Factor 4: capability
Agent work is demanding, so capability matters more here than in simple chat. Do not assume either side wins for your task.
Board results change quickly, so we will not state them here. Boards such as the Vals Index include an open-weights view, and our leaderboard has an open-weight filter. Check those for the current gap, then test on your own tasks.
Remember that a benchmark setup may give the model more resources than you would. A smaller or more compressed version on your own server may perform differently. Our methodology page explains how we choose configurations.
Factor 5: control and continuity
Proprietary vendors can retire models, change prices or change behavior, and you do not control the timing. If you need one exact version for years, that is a real risk.
A downloaded model file stays the same, and you can keep it as long as the license allows. The trade-off is that you carry its problems too, including security patches for the serving software and any flaw in the model itself. NTIA's report also discusses risks from widely available weights, and that includes lack of clear accountability.
Which should you choose?
| If this describes you | Start with |
|---|---|
| Small team, no ML engineers, normal business data | Proprietary or hosted |
| Want low token prices, and no strict data limits | Hosted open-weight, after checking the provider |
| Contract or policy forbids sending data to third parties | Self-hosted open-weight, with professional review |
| Very high steady volume, plus engineering staff | Price out self-hosting |
| Need heavy fine-tuning on your own data | Open-weight, with a permissive license |
| Not sure yet | Hosted now, with a harness that can switch models later |
That last row is the one we would give most clients. Build so that the model can be replaced, using standard connections such as MCP and our guide to remote vs local MCP. Then you can move as prices and capabilities change.
Frequently asked questions
Is open-weight the same as open source?
No. The Open Source Initiative says open weights lack the training code and dataset that would allow full study and rebuilding. Many models people call open source are more accurately open weight.
Are open-weight models free for commercial use?
Often, but not always without conditions. Llama 4's license has attribution rules and a 700 million monthly user threshold, while others use Apache 2.0. Check the license for the exact model you plan to use.
Is self-hosting more secure?
It can give you more control, since data stays on your systems. But you then own patching, access control and monitoring. A poorly run private server can be riskier than a well-run commercial service.
Can open-weight models power serious agents?
Yes, several are tested on the same agent boards as proprietary models. See the Vals Index and our board for current standing. Then test on your tasks, since a benchmark setup can differ from yours.
