Open vs. Closed LLMs Explained: How to Choose Between Open-Weight and Proprietary AI Models
· 17 min read
Your engineering lead wants to run an open model on your own servers so customer data never leaves your cloud account. Your CFO has read that open models are "free." Your security team worries that anyone can download them. Your product team just wants the model that writes the best answers. Meanwhile, the AI vendor you already pay promises it does not train on your data. Who is right?
All of them, partly. The choice between open-weight and closed large language models is one of the most practical AI decisions a business makes, and it is often framed badly, as a fight between "open source" and "Big Tech." In reality it is a set of trade-offs about control, cost, quality, licensing, security, and who carries the operational work.
This guide explains open and closed LLMs in plain English for U.S. B2B SaaS teams: what "open" actually means, how licenses differ, what self-hosting really costs, where each approach is stronger, the security and legal questions to ask, and what the spread of open models means for how AI systems describe your brand. If you are new to the technology, start with our practical guide to large language models, then come back.
Published November 17, 2026. This is an educational explainer, not legal advice. Model names, licenses, capabilities, and prices change frequently. Where this guide describes a specific model or license, it relies on the developer's public announcement or license text; check the current version of any license before you rely on it.
What "Open" and "Closed" Actually Mean
A large language model is, at its core, a very large file of numbers called weights or parameters, plus the code that runs them. Our guide to how LLMs are trained explains where those numbers come from. Whether a model is "open" depends mostly on what happens to that file.
- Closed (proprietary) models. The developer keeps the weights private. You use the model through the developer's app or API, on the developer's servers, under its terms. Examples include the main models behind ChatGPT, Claude, and Gemini.
- Open-weight models. The developer publishes the weights so anyone can download and run the model on their own hardware, usually under a license that sets conditions. Examples include Meta's Llama family, Mistral's open models, Google's Gemma, Alibaba's Qwen, DeepSeek's models, and OpenAI's gpt-oss models.
- Fully open models. Beyond weights, the developer also releases the training code, details of the training data, and often the data itself, so others can study or reproduce how the model was built. The Allen Institute for AI's OLMo project, released in 2024 with its weights, training code, and training dataset, is a well-known example.
The word "open source" causes the most confusion. In software, open source has a specific meaning, maintained by the Open Source Initiative. In October 2024 the OSI published its Open Source AI Definition 1.0, which says an open-source AI system must make available the weights, the code used to train and run it, and enough detailed information about the training data for a skilled person to build a substantially equivalent system, all under terms that allow use, study, modification, and sharing for any purpose.
Many popular "open" models do not meet that definition, because their licenses restrict some uses or they do not disclose training data in detail. That is why careful writers say open-weight rather than open source. It is not a criticism; it is simply a more accurate label for what you get.
A Short History of Open Models
Openly released language models are not new. Google published BERT in 2018 and OpenAI released the full weights of GPT-2 in 2019. What changed is that open models became large and capable enough to compete with the best commercial systems on many everyday tasks.
A few milestones explain how we got here:
- BLOOM (2022). The BigScience research collaboration, coordinated by Hugging Face with more than a thousand researchers, released BLOOM, a 176-billion-parameter multilingual model, under a "Responsible AI License" that allowed broad use but prohibited listed harmful uses.
- Llama and Llama 2 (2023). Meta first shared its LLaMA research models with approved researchers, and the weights soon circulated widely. In July 2023 it released Llama 2 for research and commercial use under its own community license, a turning point for companies that wanted a capable model they could run themselves.
- Mistral 7B (2023). French startup Mistral AI released Mistral 7B under the permissive Apache 2.0 license and reported that it outperformed larger Llama 2 models on many benchmarks, showing that smaller, well-trained models could punch above their size.
- Large open models (2024). Meta's Llama 3 herd of models paper described a 405-billion-parameter model released openly in July 2024. Google released its Gemma family, and Alibaba expanded its Qwen family, many under Apache 2.0.
- Open reasoning models (2025). In January 2025 DeepSeek released DeepSeek-R1, a reasoning model, under the MIT license. Our guide to reasoning models explains what that category does.
- Closed labs release open weights (2025). In August 2025 OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0, its first open-weight language models since GPT-2. OpenAI said the larger model could run on a single 80 GB GPU and the smaller one on devices with 16 GB of memory.
The pattern is clear: the line between "open companies" and "closed companies" has blurred. Several developers now publish some models openly while keeping their most capable models closed.
How Open Model Licenses Differ
When a model is called open, the license decides what you can actually do with it. Licenses fall into rough groups:
- Permissive licenses, such as Apache 2.0 and MIT, allow commercial use, modification, and redistribution with few conditions, mainly keeping copyright and license notices. Apache 2.0 also includes an explicit patent license.
- Custom community licenses, such as Meta's Llama licenses and Google's Gemma terms, allow broad commercial use but add conditions. The Llama 2 license, for example, required companies whose products had more than 700 million monthly active users on the release date to request a separate license from Meta, and it is paired with an acceptable use policy that prohibits certain uses.
- Responsible AI licenses, such as the one used for BLOOM, allow use but attach a list of prohibited purposes that carries over to anyone who builds on the model.
- Research or non-commercial licenses allow study and experimentation but not production business use.
For a SaaS company, the practical questions are simple. Can we use it commercially? Can we fine-tune it and ship the result to customers? Do the conditions flow down to our customers? Are there user-count thresholds, naming requirements, or acceptable use rules we must enforce? Do we need to attribute the model in our product? Have a lawyer read the actual license, because two models that are both called "open" can answer these questions very differently.
Licenses can also change between versions. A model family may move from a restrictive license to a permissive one, or the reverse, so record exactly which version and license you adopted.
What Closed Models Offer
Closed models are not simply the "locked" option. For many teams they are the sensible default.
- Frontier capability. The most capable models on difficult reasoning, coding, and multimodal tasks have usually been closed, although the gap on many everyday tasks has narrowed. Our guide to evaluating LLMs explains why you should test on your own work rather than trust leaderboards.
- No infrastructure. You send a request and get an answer. There are no GPUs to buy, no servers to scale, and no model updates to install.
- Built-in features. Commercial APIs often bundle web search, file handling, tool calling, safety filters, usage dashboards, and enterprise controls such as single sign-on and audit logs.
- Contracts and support. You get terms of service, data processing agreements, service-level commitments on some plans, and someone to call when something breaks.
- Data commitments. Major providers state that, by default, data sent through their business and API products is not used to train their models. These commitments vary by plan and change over time, so read the specific terms you sign.
The trade-offs are real too. You depend on one vendor's pricing, availability, and policy decisions. Models can be updated or retired on the vendor's schedule, which can change your product's behavior. Your data leaves your environment, even if it is handled under strong contractual terms. And you cannot inspect or modify the model itself.
What Open-Weight Models Offer
Open-weight models give you control in exchange for responsibility.
- Data stays where you put it. If you run the model on your own servers or in your own cloud account, prompts and documents never go to the model developer. For regulated data or strict customer contracts, this can be decisive.
- Stability. A downloaded model does not change unless you change it. The same weights produce the same kind of behavior next year, which matters for audits, regulated workflows, and products that must behave consistently.
- Customization. You can fine-tune the model on your own examples, compress it to run on smaller hardware, or combine it with your own retrieval system. Our guide to retrieval-augmented generation explains how that works.
- Flexible deployment. Open models can run in your data center, in any cloud, on laptops, and in some cases on phones or edge devices. Our guide to setting up an LLM locally walks through the basics.
- No per-request vendor fees. You pay for hardware and operations instead of per token. At high, steady volume this can be cheaper; at low or uneven volume it usually is not.
- Transparency. You can inspect the model's files, test it without rate limits, and in fully open cases study the training data. Stanford's Foundation Model Transparency Index, published in 2023, found that no major developer scored highly on transparency overall, while developers of open models generally scored higher than closed ones.
The cost is that you now own everything the API provider used to handle: hardware, scaling, uptime, security patching, safety filtering, monitoring, and upgrades.
What Running an Open Model Really Costs
"Free to download" is not "free to run." A useful rule of thumb: a model's weights need roughly two bytes of memory per parameter at standard 16-bit precision. A 70-billion-parameter model therefore needs about 140 GB of GPU memory just to load, before you add memory for the conversation itself. That usually means several data-center GPUs.
Quantization reduces this by storing each number with fewer bits. At 4-bit precision, the same 70-billion-parameter model needs roughly 35 GB for its weights, small enough for a single high-end GPU, usually with some loss in quality that you should measure on your own tasks. Smaller models in the 7-to-20-billion-parameter range can run on a well-equipped laptop or a single modest GPU.
Beyond memory, plan for:
- Compute for speed. Serving many users at once requires enough GPUs to keep responses fast. Serving software such as vLLM and llama.cpp helps use hardware efficiently.
- Engineering time. Someone has to deploy, monitor, update, and secure the system. For a small team, that salary often outweighs any savings on API fees.
- Idle capacity. GPUs cost money whether or not anyone is using them. APIs charge only for what you use.
- Context costs. Long documents and conversations increase memory and compute needs. Our guide to tokens and context windows explains why.
A practical approach is to estimate your monthly volume in tokens, price it on two or three commercial APIs, then price the hardware and staff time to serve the same volume yourself at your required speed. Many teams find APIs cheaper until volume is large and predictable. You can also rent open models from cloud providers that host them for you, which keeps the model choice open without owning the servers.
Quality: Is the Gap Closing?
On many common business tasks, such as summarizing documents, classifying support tickets, extracting fields, drafting routine emails, and answering questions from a provided knowledge base, current open-weight models perform well. On the hardest tasks, such as long multi-step reasoning, complex coding, and advanced multimodal understanding, the strongest closed models have generally led, although open releases have repeatedly narrowed the gap within months.
Three practical points matter more than leaderboard rankings:
- Task fit beats general rank. A smaller model fine-tuned on your exact task can outperform a much larger general model on that task, at a fraction of the cost.
- The system matters as much as the model. Good retrieval, clear prompts, and well-designed tools often improve results more than switching models. Our guide to prompt engineering covers the basics.
- Measure on your own examples. Build 50 to 100 real examples with known good answers, run each candidate model, and compare accuracy, cost, and speed. Our guide to LLM hallucinations explains why checking factual accuracy matters for every model, open or closed.
Security and Safety Questions
Both approaches carry security risks, but they sit in different places. Our guide to LLM security and privacy covers the general threats, such as prompt injection and data leakage, which apply to both.
With closed models, your main concerns are what data you send, how the provider stores and processes it, who on your team can access the account, and how you handle provider outages or policy changes.
With open-weight models, you take on extra responsibilities:
- Supply chain. Download models only from the developer's official repository, verify file checksums where provided, and prefer the safetensors file format, which was designed to avoid the risk that older Python "pickle" files can run hidden code when loaded.
- Safety behavior. Closed APIs add their own filters; a self-hosted model has only the safety training built into it, plus whatever you add. Research such as Qi et al. (2023) showed that even small amounts of fine-tuning can weaken a model's safety behavior, so test again after any customization.
- Patching and monitoring. You are responsible for keeping the serving software updated, restricting network access, logging use, and watching for abuse.
- Agents and tools. If an open model can call tools or take actions, the same guardrails apply as with any model. Our guide to LLM agents and tool use explains why human approval for sensitive actions matters.
At the policy level, open weights raise a debate about misuse: once weights are published, they cannot be recalled. In July 2024 the U.S. National Telecommunications and Information Administration published a report on dual-use foundation models with widely available model weights, which concluded that the government should not immediately restrict open weights but should build the capacity to monitor risks and act if needed. State laws are also emerging; our guide to California's SB 53 explains obligations for the largest frontier model developers. Most companies that simply use open models are not frontier developers, but your counsel should confirm what applies to you.
Legal and Compliance Questions to Ask
Before adopting any model, open or closed, work through a short checklist with legal and security:
- License or terms. Does it permit our intended commercial use, fine-tuning, and redistribution? What conditions flow down to our customers?
- Acceptable use. Which uses are prohibited, and how will we enforce that in our product?
- Data handling. For closed models, what are retention, training, and processing terms on our plan? For open models, where will data be stored and who can access the servers?
- Training data and copyright. What does the developer disclose about training data, and does the provider offer any legal protection to customers? Commitments differ widely between providers and plans.
- Customer contracts. Do our customer agreements, security questionnaires, or data processing addendums restrict which AI vendors or regions we can use? Our article on how AI engines read SaaS data policies explains why buyers increasingly ask.
- Documentation. Can we record which model version, license, and configuration produced a given output, for audits and incident response?
How to Choose: A Practical Framework
Most teams do not need a single answer. A common pattern is a closed model for difficult, low-volume tasks and an open model for high-volume, narrow, or sensitive tasks. Use these questions to decide each workload:
- Choose a closed model when you need the highest capability, have a small team, have low or unpredictable volume, need built-in search or multimodal features, or want contractual support and fast setup.
- Choose an open-weight model when data must stay in your environment, behavior must stay stable for long periods, volume is high and predictable, you need deep customization or offline use, or you want to avoid depending on one vendor.
- Consider a hosted open model when you want open-model flexibility and portability without operating GPUs yourself.
Whatever you choose, design for change. Keep prompts, evaluation sets, and retrieval separate from any single model, so you can switch when a better or cheaper option appears. In this market, that happens often.
What Open Models Mean for Your Brand's Visibility
For B2B SaaS teams, there is a second reason to understand open models: they shape how AI describes your company in places you never see.
Open models are everywhere, and many are frozen
Open-weight models are built into customer support bots, internal knowledge tools, browser extensions, coding assistants, and niche AI products across the industry. Each copy carries the knowledge it absorbed during training, up to its training cutoff, and keeps it until whoever runs it updates to a newer model. Some products add live search or retrieval; many do not. That means an outdated description of your product can keep circulating long after you have changed your pricing, positioning, or features.
You cannot submit a correction to a model that has already been downloaded. What you can do is make sure the public record that future models learn from, and that retrieval systems fetch, is clear, consistent, and dated.
Consistency across the public record matters
Because no single model or vendor controls what all these systems learn, the most durable signal is a consistent public description of your company across many sources: your own site, documentation, reputable industry publications, partner pages, and community discussions. If your homepage, pricing page, and third-party listings disagree on what you do or who you serve, models trained on that mix may repeat any version. Our guide to structuring a website for LLM crawlers covers the basics of making your facts easy to read.
Open models make observation easier to repeat
Open-weight models have one advantage for anyone measuring AI visibility: a downloaded model does not change silently. If you record the exact model version, prompt, settings, and date, you can rerun the same test later and compare. Consumer AI apps, by contrast, can change models, add search, or personalize answers without notice. Observing both is useful: consumer engines show what buyers see today, and fixed open models show what a model learned about you from its training data. Our guide to tracking LLM recommendations explains how to record observations properly.
None of this guarantees that any model will mention or recommend your brand. It removes avoidable reasons for AI systems to describe you incorrectly, and it gives you a repeatable way to check what they actually say.
A 30-Day Plan for SaaS Teams
Week 1: Map your AI workloads. List every place your company uses or plans to use an LLM: support, sales, internal search, product features. For each, note the data involved, monthly volume, quality needs, and any customer contract limits.
Week 2: Build an evaluation set. Collect 50 to 100 real examples for your top two workloads, with agreed good answers. Run one or two closed models and one or two open-weight models, and compare accuracy, cost per task, and speed.
Week 3: Review licenses and terms. Have legal review the license or terms for each candidate, and security review data handling, supply chain, and access controls. Record the exact model versions you tested.
Week 4: Check how AI describes you. Run a short, fixed set of buyer questions about your category on two consumer AI engines and on one open-weight model running locally. Record the date, model, prompt, and exact answer. Note any outdated or wrong facts, and fix the public pages they likely came from.
The Bottom Line
Closed models give you top capability and convenience in exchange for dependence on a vendor. Open-weight models give you control, stability, and portability in exchange for running and securing them yourself. "Open" covers everything from permissive Apache 2.0 releases to custom licenses with real conditions, so read the license, not the label. For most businesses, the right answer is a mix, chosen workload by workload and tested on real examples.
For brands, the spread of open models means AI descriptions of your company live in many places you cannot update directly. A clear, consistent, dated public record is the best defense. If you want to see what AI engines currently say about your brand, and which sources they rely on, book a strategy call with LLM Recommend.
Frequently Asked Questions
What is the difference between open and closed LLMs?
A closed LLM keeps its weights private, so you can only use it through the developer's app or API. An open-weight LLM publishes its weights so anyone can download and run it on their own hardware, usually under a license with conditions. Fully open models also share training code and training data details.
Are open-weight models the same as open-source models?
Not always. The Open Source Initiative's Open Source AI Definition requires the weights, the training and inference code, and detailed training data information under terms allowing any use. Many popular open-weight models restrict some uses or do not disclose training data in detail, so they are more accurately called open-weight.
Are open-source LLMs free to use?
The download is usually free, but running the model is not. You pay for GPUs or cloud hosting, engineering time, monitoring, and security. A 70-billion-parameter model needs about 140 GB of memory at 16-bit precision. At low volume, a paid API is often cheaper than self-hosting.
Are open-weight models less secure than closed models?
Not inherently, but the responsibilities differ. Self-hosting keeps data in your environment, but you must verify model files, add safety filters, patch software, and monitor use. Closed APIs handle much of that, but your data leaves your environment under the provider's terms. Both are exposed to prompt injection.
Can my company use open models commercially?
Often, yes, but it depends on the license. Apache 2.0 and MIT models allow broad commercial use. Custom licenses, such as Meta's Llama licenses, add conditions like user-count thresholds and acceptable use policies. Some models allow research use only. Have counsel review the exact license version.
How do open models affect my brand's visibility in AI answers?
Open models are built into many products and keep the knowledge from their training data until they are replaced, so outdated facts about your brand can persist. Keep your public description clear, consistent, and dated across your site and trusted third-party sources, and observe both consumer engines and fixed open models over time. This reduces errors but never guarantees inclusion.