Open Source AI vs Proprietary AI: The Real Tradeoffs

Meta calls Llama "open source." The Open Source Initiative, the nonprofit that has stewarded the definition of that term for software since 1998, disagrees. The dispute is not academic. It decides whether a hospital in Germany can deploy a model without a compliance review, whether a startup can fine-tune it and resell the result, and whether a researcher can ever see the data the model learned from. This article separates the marketing from the license terms and weighs open models against closed ones like GPT-5 and Claude on cost, control, capability, and law.

What "open" actually means, and why it's contested

Most AI companies that describe their models as open source are really shipping open weights. Open weights means you get the finished, trained model file: the billions of numerical parameters you can download and run yourself. Think of it as being handed a baked cake. Open source, in the traditional software sense, means you also get the recipe: the code that trained the model, documentation of the data, and enough detail to bake it again and audit how it was made.

The Open Source Initiative (OSI) made that distinction formal. Its Open Source AI Definition, version 1.0, was announced in late October 2024. It says an open source AI system must grant four freedoms: to use the system for any purpose without asking permission, to study how it works and inspect its components, to modify it for any purpose, and to share it with or without modifications. To make those freedoms real, the definition requires three things in the "preferred form for making modifications": sufficiently detailed information about the training data, the complete source code used to train and run the system, and the model parameters (Open Source Initiative).

Not everyone accepted the compromise. The definition asks for "data information" rather than the data itself, and Bradley Kuhn argued that the definition sets too few requirements on training-data disclosure for the public to reproduce how these systems are built, and that it therefore erodes the meaning of "open source" (Kuhn, October 31, 2024). Even under the lenient reading, many popular "open" models fail on at least one requirement, usually the training code or the data description.

How Llama's license differs from a true open source license

Meta's Llama 4 Community License shows how a model can be free to download and still carry strings. Four clauses matter most:

• A user-count gate. If the licensee's products had more than 700 million monthly active users in the month before the release date, the licensee must request a license from Meta, which Meta "may grant to you in its sole discretion."
• An incorporated usage policy. Users must adhere to Meta's Acceptable Use Policy, which the license incorporates by reference. That is a field-of-use restriction by design.
• A naming rule. If you use Llama materials or their outputs to create, train, fine-tune, or otherwise improve an AI model that you distribute, the new model's name must begin with "Llama."
• Distribution terms. Redistributors have to pass on the license and the policy.

All four are visible in the license text published with the April 5, 2025 release (Llama 4 license text, mirrored by Wind River). The OSI's February 18, 2025 post, "Meta's LLaMa license is still not Open Source," argues that this style of license fails the Open Source Definition because it discriminates among users, restricts fields of endeavor, and, in newer versions, excludes people located in the European Union from some uses (OSI).

One correction to a common claim: older discussions say the Llama license bans using outputs to train any other model. The Llama 4 text instead conditions that use on the naming rule above. Licenses change between versions, so the version you actually download is the one that binds you.

Which "open" models use permissive licenses

Other vendors have moved toward standard licenses that carry none of those strings:

Model family	License on weights	Notable terms
DeepSeek V3 and V4	MIT	Permissive; no user-count gate
Qwen3 (for example Qwen3-235B-A22B)	Apache 2.0	Permissive; patent grant included
Mistral Large 3 and Ministral 3	Apache 2.0	Released December 2, 2025
OpenAI gpt-oss-120b	Apache 2.0	Released August 2025; runs on a single 80 GB GPU
Llama 4	Llama 4 Community License	700 million user gate, usage policy, naming rule
Gemma (earlier releases)	Gemma Terms of Use	Prohibited Use Policy; Google may restrict use it believes violates the terms

Sources: DeepSeek-V4-Pro model card, Qwen3-235B-A22B model card, Mistral 3 announcement, gpt-oss-120b model card, Gemma Terms of Use.

Even a fully permissive license on the weights does not make the project open source in the OSI sense. None of these vendors publishes its complete training data, and some publish little training code. They are open weight with a good license, which is a big step up from open weight with a restrictive one.

The performance gap between the best closed model and the best open-weight model on the Chatbot Arena leaderboard shrank from 8.04 percent in early January 2024 to 1.70 percent by February 2025, according to Stanford HAI's AI Index Report 2025 ([Stanford HAI](https://hai.stanford.edu/ai-index/2025-ai-index-report/technical-performance)).

Where each model type actually wins

Proprietary models: the hosted-service advantage

Closed models such as GPT-5, Claude, and Gemini are sold as a service, not a product you install. You send a request to an API (an application programming interface, a standard way for software to call another service) and get a response back. The company handles the GPUs, the uptime, the safety filtering, and the model updates. That convenience is the entire value proposition, in the same way a taxi beats owning a car if you only need a ride occasionally: you pay per use and never touch the engine.

The model also concentrates capability. Frontier labs spend enormous compute budgets on a small number of flagship releases, and historically that has kept the very top of the benchmark leaderboards closed. Proprietary vendors also bundle content moderation, abuse monitoring, compliance paperwork, and enterprise support contracts that most self-hosted deployments have to build or buy separately.

There is a quieter advantage as well: you inherit improvement without effort. When a vendor ships a better model behind the same endpoint, your application gets the upgrade the next time you change a version string. A self-hosted team has to evaluate, download, re-quantize, and re-test every new release itself.

Open weight models: the control-and-customization advantage

Running an open weight model is closer to owning a car than renting a taxi. You can modify the engine, drive anywhere without asking permission, and nobody can revoke your access because you exceeded a quota. Once the weights are on your disk, no API outage, price hike, or policy change can take the model away. If a vendor deprecates a closed model, as vendors routinely do, every product built on it must migrate. A downloaded checkpoint stays exactly as it was.

That control has concrete engineering payoffs. A company handling regulated financial or medical records can run an open weight model entirely inside its own network, so no prompt or document ever crosses a vendor boundary, which is often a hard requirement under data-residency rules. Teams can also fine-tune open weights on proprietary data to build a narrow specialist, something many closed-model terms restrict or price heavily.

Open weights also make research reproducible. A paper that evaluates a closed model is, in a strict sense, evaluating a moving target that may be silently updated. A paper that evaluates a pinned open checkpoint can be rerun by anyone, which is why academic work on interpretability and safety leans so heavily on open models. For a look at what that interpretability work involves, see mechanistic interpretability.

Is open source AI cheaper than proprietary AI?

Not automatically, and never free. Open weights remove the per-token API fee, but you still pay for GPUs, storage, networking, monitoring, and engineers. For steady, high-volume use, self-hosting can be cheaper over time. For sporadic or low-volume use, an API's pay-per-call pricing is usually cheaper because you never buy idle hardware.

A worked example makes the break-even logic concrete. The numbers below are hypothetical, chosen only to show the arithmetic:

• Suppose renting a GPU server costs $8 per hour. Running it around the clock for 30 days costs $8 x 24 x 30 = $5,760 per month.
• Suppose a comparable hosted API charges a blended $2 per million tokens (input and output averaged). Then $5,760 buys 2,880 million tokens, or 2.88 billion tokens.
• To match that spend, your server has to process 2.88 billion tokens in 2,592,000 seconds, an average of about 1,111 tokens per second, every second of the month.

If your real traffic averages 100 tokens per second, the hosted API is roughly an order of magnitude cheaper, because the server sits idle most of the day. If you can keep the hardware saturated, the self-hosted route starts to win, and techniques like quantization and batching engines such as vLLM stretch each GPU further. Engineering salary is the cost most spreadsheets forget, and at small scale it dwarfs the hardware bill.

The capability gap has narrowed, not closed

For most of 2023 and early 2024, the story was simple: closed frontier models were meaningfully smarter than anything you could download. That gap has compressed quickly. Stanford's AI Index measured it on one widely watched leaderboard, Chatbot Arena, where human voters compare model answers side by side. The distance between the top closed and top open-weight model fell from 8.04 percent in January 2024 to 1.70 percent in February 2025.

The 2026 AI Index reports a related result across national lines. It says the gap between US and Chinese models has "effectively closed," with the top US model leading by 2.7 percent as of March 2026 (an Arena score of 1,503 versus 1,464) (Stanford HAI, 2026). That is a statement about countries, not about licensing, but several leading Chinese models are open weight, so the two trends are linked. Note, too, that leaderboard gaps of this size sit close to the noise level of human-preference voting.

What the big open releases look like in numbers

The recent open releases show how large the frontier of downloadable models has become:

Model	Released	Total parameters	Active per token	Context window	Source
Llama 4 Scout	April 5, 2025	109 billion	17 billion	10 million tokens	Meta
Llama 4 Maverick	April 5, 2025	400 billion	17 billion	Not stated in the announcement	Meta
DeepSeek-V3	December 2024	671 billion	37 billion	Not listed here	arXiv:2412.19437
Qwen3-235B-A22B	2025	235 billion	22 billion	32,768 native; 131,072 with YaRN scaling	Hugging Face
gpt-oss-120b	August 2025	117 billion	5.1 billion	Not listed here	Hugging Face
Mistral Large 3	December 2, 2025	675 billion	41 billion	Not listed here	Mistral
DeepSeek-V4-Pro (preview)	April 2026	1.6 trillion	49 billion	1 million tokens	Hugging Face

Every row uses a mixture-of-experts design. Instead of activating every parameter for every word, the model routes each token to a small subset of specialist sub-networks. It is like a hospital where a triage nurse sends each patient to the right specialist rather than making every doctor examine everyone. You get the knowledge capacity of a very large model at the running cost of a much smaller one. The mixture-of-experts explainer covers the routing mechanism in detail.

One practical consequence is easy to overlook. Active parameters set the compute per token, but total parameters set the memory you need, because all the experts must be loaded. At 8-bit precision, 1.6 trillion parameters is roughly 1.6 terabytes of weights before any working memory, far beyond a single GPU. Meta says Llama 4 Scout, at 109 billion parameters, fits on a single NVIDIA H100 GPU with Int4 quantization, and the gpt-oss-120b model card says that model is designed to fit on one 80 GB GPU. "Open" does not mean "runs on your laptop." For models that do, see open source AI models you can run locally and the hidden costs of running a local LLM.

Why leaderboard numbers deserve suspicion

Benchmarks are easy to game and hard to interpret. The Llama 4 launch is the clearest recent case. Meta's announcement said Maverick beat GPT-4o and Gemini 2.0 Flash "across a broad range of widely reported benchmarks" and matched DeepSeek v3 on reasoning and coding with less than half the active parameters (Meta). But the version of Maverick that scored highly on LMArena, labeled "Llama-4-Maverick-03-26-Experimental," was an experimental chat variant "optimized for conversationality," not the model Meta released. When LMArena scored the unmodified public model, it ranked below GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro, and LMArena apologized and changed its policies (TechCrunch, April 11, 2025).

The lesson applies to closed vendors too: a vendor-reported score tells you what the vendor chose to measure. Run your own evaluation on your own tasks before committing, and read why AI benchmarks can be misleading for the common failure modes.

Why did open weight models catch up so fast?

Three forces compressed the gap. First, distillation: training a smaller model on the outputs of a stronger one transfers capability cheaply. Second, mixture-of-experts designs let labs scale capability without scaling inference cost in proportion. Third, open research is cumulative: when DeepSeek-R1 published its reasoning-training recipe, other labs could adopt the technique within weeks instead of re-inventing it.

The licenses and the law: what "open" buys you in the EU

The EU AI Act treats "open source" as a legal category with real consequences, which raises the stakes of the labeling fight. Two separate provisions are often confused.

The first is Article 2(12), which says the Act does not apply to AI systems released under free and open source licenses. The exemption has exceptions: it does not cover systems placed on the market as high-risk, systems that fall under the prohibited practices of Article 5, or systems with transparency duties under Article 50, such as those that interact directly with people or generate content (Linux Foundation Europe).

The second concerns general-purpose AI (GPAI) models, the category that includes large language models. According to the European Commission, obligations for GPAI providers entered into application on August 2, 2025, and from August 2, 2026 the Commission will enforce full compliance, including through fines. A model trained with more than 10^23 floating-point operations (FLOP, a count of arithmetic steps) that can generate language is treated as GPAI under an "indicative criterion," and a cumulative training compute above 10^25 FLOP creates a presumption of systemic risk (European Commission FAQ).

Open source providers can skip some paperwork, namely documentation for authorities, documentation for downstream providers, and appointing an EU representative, but only if all three conditions hold:

1. The model is under a truly free and open-source license that allows access, use, modification, and distribution without monetization.
2. Its parameters, including weights, architecture, and usage information, are publicly available.
3. It is not classified as a GPAI model with systemic risk.

Even then, open source providers must still maintain a copyright policy and publish a summary of training data (European Commission FAQ).

Does the EU AI Act treat open weight models the same as open source?

No. The exemption is written around free and open-source licensing criteria, not around the looser "open weight" label. A license with a user-count gate and a field-of-use policy, like Llama's, does not clearly meet the "truly free" bar, and the exemption disappears for models posing systemic risk regardless of license. Simon Willison, a software developer who writes widely on AI, floated the theory that Meta's insistence on the "open source" label is partly motivated by these rules, while noting he is not a lawyer and that it is his reading rather than a confirmed explanation (Simon Willison, April 19, 2025).

For a company, the practical point is simpler than the legal theory. Compliance duties attach to who you are and what you do with a model, not to the brand on the download page, so legal review should read the license and the Act together.

What people get wrong about cost and safety

The most persistent myth is that open models prove frontier AI is cheap. DeepSeek's V3 technical report said the model's "total training costs amount to only $5.576M," computed from 2.788 million H800 GPU hours at an assumed rental price of $2 per GPU hour (2,788,000 x $2 = $5,576,000). The same report states the figure covers only the official training of V3 and excludes "the costs associated with prior research and ablation experiments on architectures, algorithms, or data" (DeepSeek-V3 Technical Report).

Analysts took the figure apart further. SemiAnalysis called the headline "$6M" number misleading as a measure of what DeepSeek spends, because it omits research, experimentation, and the total cost of owning the hardware. It estimated DeepSeek's total server capital expenditure at about $1.6 billion, with roughly $944 million tied to operating the clusters, and believed the company had around 50,000 Hopper-generation GPUs, a mix of H800, H100, and H20 chips (SemiAnalysis). These are estimates from an analyst firm, not audited figures. The training run was cheap once the lab existed; building the lab was not.

The safety argument runs in both directions

A second myth says closed models are inherently safer because a company controls them. Centralized control does make some things easier: a vendor can patch a model, rate-limit abusers, and pull a capability after release. But control also means every user depends on one company's judgment about what counts as harmful, and outsiders cannot audit that judgment.

Open weights flip the tradeoff. Any lab or researcher can red-team the model and publish findings, which spreads scrutiny. But anyone can also fine-tune the safety training away, because the weights are already on their machine. Neither approach removes risk. They distribute it differently, and which distribution is preferable is a live policy debate among safety researchers, not a settled answer. The AI safety and alignment explainer lays out the main positions.

"Open" is not one tier

A third confusion treats open as a single category. Llama's community license, Apache 2.0 releases from Mistral and Qwen, MIT-licensed DeepSeek weights, and Google's Gemma terms (which include a Prohibited Use Policy, flow-down requirements for redistributors, and a clause letting Google restrict use it reasonably believes violates the terms) carry different obligations. Gemma's terms even define "Model Derivatives" to include models trained to mimic Gemma through distillation or synthetic data generated by Gemma (Gemma Terms of Use). Before building a product on any of them, read the actual license for the exact version you download, not the announcement page.

A decision framework: which approach fits which situation

The choice is rarely binary. Many mature teams run both: a closed API for the hardest reasoning tasks and a self-hosted open model for high-volume, privacy-sensitive, or latency-critical work. Services that route each request to the cheapest model able to handle it (model routing) formalize that split.

Situation	Usually leans toward	Why
Prototype or low, unpredictable traffic	Proprietary API	No idle hardware; fastest start
Data may not leave your network	Open weight, self-hosted	Prompts never cross a vendor boundary
Sustained high volume with steady load	Open weight, self-hosted (verify with your own math)	Fixed hardware cost spreads over many tokens
Need the strongest reasoning available today	Proprietary API, then test open models	Frontier lead is narrow but real on hard tasks
Need to fine-tune on proprietary data	Open weight with permissive license	Full control of weights and training
Need uptime guarantees and a vendor to call	Proprietary API	Service-level agreements and enterprise support
Product distributed to millions of users	Read the license first	Llama's 700 million user gate matters only at huge scale, but MIT and Apache 2.0 avoid the question
Regulated industry needing audit trails	Either, after legal review	EU AI Act duties depend on use and risk class

Two further points help in practice. First, switching costs are lower than they look, because most inference servers and vendors expose OpenAI-compatible endpoints, so an application can often swap models by changing a base URL and a model name. Second, the smaller end of the open ecosystem (models in the 3 billion to 14 billion parameter range, such as Mistral's Ministral 3 family) is where open models are most clearly the right default, since a small, cheap specialist often beats a giant generalist on a narrow job. See what small language models are good for.

Frequently Asked Questions

Q: What is the main difference between open source AI and proprietary AI? A: Open weight and open source models can be downloaded, inspected to varying degrees, and modified under their license terms. Proprietary models are reachable only through a company's API or app, with no access to the weights. Strictly, the OSI's definition of open source AI also requires training code and data information, which most "open" models do not provide.

Q: Is Llama actually open source? A: Not by the OSI's definition. The Llama 4 Community License requires companies above 700 million monthly active users to request a license that Meta may refuse, incorporates an Acceptable Use Policy, and requires a "Llama" name prefix on models trained on its outputs. The OSI argued in February 2025 that Llama licenses fail the Open Source Definition.

Q: Are open weight models as capable as proprietary models? A: They are close on many benchmarks but usually still slightly behind at the very top. Stanford's AI Index found the closed-versus-open gap on Chatbot Arena fell from 8.04 percent in January 2024 to 1.70 percent in February 2025. The 2026 AI Index reports that the US-China model gap has effectively closed, at 2.7 percent as of March 2026.

Q: Is open weight AI cheaper to run than a proprietary API? A: It depends on volume. An API has no upfront cost and is usually cheaper for low or irregular traffic. Self-hosting means paying for GPUs and engineers whether or not requests arrive, so it tends to win only when hardware stays busy. Compute the break-even with your own token counts and prices before deciding.

Q: Does the EU AI Act exempt open source models? A: Partly. Article 2(12) exempts AI systems under free and open source licenses unless they are high-risk, prohibited, or subject to Article 50 transparency duties. For general-purpose models, open source providers can skip some documentation duties if the license is truly free, the weights are public, and the model is not classified as systemic risk. Copyright policy and training-data summary duties still apply.

Q: Why was DeepSeek's $5.6 million training cost controversial? A: The figure in the V3 technical report covers only the official training run, priced at $2 per H800 GPU hour for 2.788 million hours, and excludes prior research and ablation experiments. SemiAnalysis estimated DeepSeek's total server capital expenditure at about $1.6 billion. The headline number measures one run, not the cost of building the lab.

Q: Which should a startup choose, open or proprietary AI? A: A startup with strict data-residency needs, high steady volume, or a need to fine-tune on private data often benefits from a self-hosted open model. A startup that needs to ship quickly, has unpredictable traffic, and wants the vendor to handle uptime and safety tooling usually does better with a proprietary API. Many teams start with an API and move specific workloads to open models later.

The takeaway

"Open" in AI marketing rarely means what "open source" has meant in software for three decades, and the gap between the label and the license text carries real legal and technical consequences. The question that matters is less about which model tops a leaderboard this month, since that gap keeps shrinking and shifting, and more about who needs control over the deployment. A team that must keep data in-house, audit behavior, or customize a model gets more from open weights than from any API, even after counting the extra infrastructure work. A team that values speed, a vendor to call, and upgrades without effort is paying for something real when it rents a closed model.

---

Read next:

• AI Agents Explained — How Autonomous AI Systems Actually Work
• GEO Explained — Generative Engine Optimization
• How to Rank in AI Overviews — AEO Guide (2026)
• Llama 4 Explained — Meta's Open AI Models
• What Is Mistral AI? Europe's LLM Lab, Explained

Sources

• Open Source Initiative — Open Source AI Definition (OSAID) v1.0
• Open Source Initiative — "Meta's LLaMa license is still not Open Source" (February 18, 2025)
• Llama 4 Community License Agreement text (April 5, 2025), mirrored by Wind River
• Meta AI — The Llama 4 herd
• TechCrunch — Meta's vanilla Maverick AI model ranks below rivals on a popular chat benchmark (April 11, 2025)
• Stanford HAI — 2025 AI Index Report, Technical Performance chapter
• Stanford HAI — 2026 AI Index Report
• DeepSeek-V3 Technical Report (arXiv:2412.19437)
• DeepSeek-V4-Pro model card, Hugging Face
• Qwen3-235B-A22B model card, Hugging Face
• Mistral AI — Introducing Mistral 3 (December 2, 2025)
• OpenAI gpt-oss-120b model card, Hugging Face
• Google — Gemma Terms of Use
• European Commission — FAQ on guidelines for obligations of general-purpose AI providers
• Linux Foundation Europe — What Open Source Developers Need to Know about the EU AI Act
• Simon Willison — "Maybe Meta's Llama claims to be open source because of the EU AI Act" (April 19, 2025)
• SemiAnalysis — DeepSeek Debates
• Bradley Kuhn — "Open Source AI Definition Erodes the Meaning of Open Source" (October 31, 2024)