Open vs closed AI models is the choice that now sits under most serious AI decisions: do you run a model you can download and control, or rent one through a vendor’s API? The honest framing has three tiers, not two. A model can be fully open source (weights, training data, and training recipe all published), open weight (the trained weights released, but the data and recipe held back), or closed and proprietary (available only as a hosted service). This piece maps the full spectrum, the license fine print, the tradeoffs on cost, control, and data governance, and the safety argument that runs underneath all of it.
The openness spectrum has three tiers, not two
Most coverage treats AI models as either "open" or "closed." That framing hides the tier that matters most in practice. Openness is a spectrum, and the middle of it is where nearly every model people call "open source" actually sits.
At one end is fully open source: the trained weights, the training data (or a detailed account of it), and the training code are all published, so a competent team could study, audit, and rebuild the model from scratch. In the middle is open weight: the finished weights are released for download and self-hosting, but the training data and the recipe stay private. At the other end is closed and proprietary: the model lives behind a vendor’s API, and you rent access rather than possess the model.
The important correction, per the Open Source Initiative, is that open weight is not the same as open source. Almost every model marketed as "open source" today, including Llama, DeepSeek, Qwen, and Gemma, is open weight only. Getting the tier right changes what you can actually do with a model.
Fully open source: weights, data, and the recipe
The strict definition arrived in October 2024, when the Open Source Initiative published the Open Source AI Definition (OSAID) 1.0. To qualify, a model has to give users enough to study, modify, and reproduce it: the weights, the training and inference code, and sufficiently detailed information about the training data for a skilled person to substantially recreate the system.
Notably, OSAID does not force publishers to ship the raw training dataset itself. The OSI made that compromise because much foundation-model data is encumbered by copyright, contracts, or privacy law, per AIwire’s reporting. What it does require is provenance and process detail, not a black box.
Very few frontier-scale models clear this bar. The ones that do tend to come from research labs that treat full reproducibility as the point, such as the Allen Institute for AI’s OLMo line and EleutherAI’s earlier Pythia work. For most businesses, fully open source is less about a specific model and more about the guarantee: nothing about how the model was made is hidden from you.
Open weight: the weights without the recipe
This is the tier doing the real competitive work in 2026, and it is dominated by Chinese labs. Open weight means the vendor publishes the trained parameters, usually on Hugging Face, so you can download the model, run it on your own hardware, and fine-tune it. What you do not get is the training data or the full recipe.
The clearest examples are the recent Chinese frontier releases. DeepSeek V4 Pro is a 1.6-trillion-parameter mixture-of-experts model released under the permissive MIT license, with weights on Hugging Face. GLM-5.2 from Zhipu AI is likewise MIT-licensed and coding-first, beating GPT-5.5 on several long-horizon coding benchmarks at a fraction of the price. Both are close to "pure open" because MIT imposes almost no restrictions.
Others show how slippery the "open" label can be, and how quickly the picture moves. Kimi K3, Moonshot AI’s 2.8-trillion-parameter model, was announced as open in mid-July but was initially reachable only through Moonshot’s own API, which is not the same as having the weights in hand. Moonshot then published the full weights on July 26, making K3 the strongest open-weight model released to date. The license is its own lesson in reading the fine print: the Kimi K3 License is derived from MIT and is permissive for almost everybody, but it attaches two conditions at scale. Run a model-as-a-service business above roughly $20 million in revenue over any twelve months and you need a separate agreement with Moonshot; ship a product with more than 100 million monthly active users and you must display "Kimi K3" in the interface. That is more open than most, and still not plain MIT. Four infrastructure providers stood K3 up on release day, and what they charge is a useful reality check on the assumption that open weights automatically mean cheaper tokens, which we broke down in our look at what Kimi K3 hosting actually costs. Qwen3.8, Alibaba’s 2.4-trillion-parameter flagship, was previewed with a promise to go open weight "soon," with no license, model card, or repository on launch day. A promise to open weights later is a roadmap, not a release, and the gap between the two is exactly where the K3 announcement sat for ten days.
Closed and proprietary: the API-only frontier
At the closed end sit the models most people use daily. Claude Opus 4.8 from Anthropic, GPT-5.6 from OpenAI, and Google’s Gemini line are all proprietary: you reach them only through a vendor API or app, and the weights never leave the vendor’s infrastructure. The current frontier set is still led by these closed models on the hardest general reasoning tasks.
Closed is a strategy, not a default, and even open-leaning labs make the choice deliberately. Meta built its reputation giving away Llama weights, but Muse Spark 1.1, its agentic flagship from Meta Superintelligence Labs, is proprietary and closed weight, sold through Meta’s first paid model API. The pattern across vendors is consistent: open the smaller or older models, keep the flagship closed.
The license spectrum, from MIT to bespoke terms
"Open" is a license question as much as an access question, and the licenses form their own spectrum.
- Permissive open-source licenses (MIT, Apache 2.0): free to use, modify, redistribute, and commercialize, with no royalty and no user cap. DeepSeek V4 Pro and GLM-5.2 use MIT; Qwen’s smaller dense models have long shipped under Apache 2.0. This is the closest thing to “no strings.”
- Custom community licenses: weights are downloadable, but the terms are the vendor’s own, not an OSI-approved license. Meta’s Llama Community License is the archetype. It requires any company above 700 million monthly active users to request a separate license Meta can refuse, per TechTarget. The OSI has been blunt that this is “not open source.”
- Closed API terms of service: no weights at all, only a contract governing usage of the hosted service.
The gap between marketing and license is wide enough to have a name. The OSI and others call the practice of claiming an open label while keeping weights, data, or code private "open washing." The practical lesson holds across the board: read the actual license, not the press release, before you build on a model.
Self-hosting, data governance, and customization
The case for open weight and open source comes down to control, and it rests on three concrete advantages.
The first is data governance. When you self-host a model, prompts and outputs never leave infrastructure you control. For regulated data, air-gapped environments, or anything under strict privacy rules, that can be the deciding factor. The second is customization: with the weights in hand, you can fine-tune on your own data and shape the model to your domain, which a closed API rarely permits at the same depth. The third is durability: an open-weight model keeps working regardless of a vendor’s business decisions, pricing changes, or deprecations. A model you have already downloaded cannot be switched off from the outside.
The costs are real too. Self-hosting a frontier-scale model is an infrastructure project. A 2.4-trillion-parameter model at 4-bit precision needs roughly 1.2 terabytes just to hold the weights, before any serving overhead, which is why raw parameter counts are a poor guide to what a team can actually run. You own the serving, the scaling, the security patching, and the uptime. Control and burden arrive together.
Convenience, cost, and the hosted-API case
The case for closed, hosted models is convenience, and for most teams it is a strong one. There is nothing to provision. You get the latest version automatically, the vendor handles scaling and reliability, and you pay only for the tokens you use.
Cost is not a clean win for either side. On hosted-API pricing, several open-weight models are dramatically cheaper: GLM-5.2 runs around $1.40 per million input tokens and $4.40 per million output, against roughly $5 and $25 for Claude Opus 4.8, or $5 and $30 for GPT-5.6 Sol. But self-hosting trades per-token pricing for fixed infrastructure cost, which only pays off at high, steady volume. At low or spiky volume, a hosted API is usually cheaper as well as simpler. That math is why many teams route by task: send the bulk of easy requests to a cheap model, and reserve a frontier model for the hard ones.
The open vs closed AI models safety debate
Underneath the practical tradeoffs runs a genuine safety disagreement, and it is unresolved. In 2024 the US National Telecommunications and Information Administration studied exactly this question for models with widely available weights.
The case against wide-open weights is proliferation. Once weights are downloadable, they can be copied and redistributed beyond the developer’s knowledge or control, and safety fine-tuning can be stripped back out. The NTIA report notes the concern that open weights could make it easier for non-experts to misuse models, including toward dangerous ends. The case for openness is that the same transparency enables safety: independent researchers can audit, red-team, find vulnerabilities, and study interpretability in ways a closed API forbids.
The NTIA’s own conclusion was cautious. It framed the question as one of marginal risk, the risk unique to open-weight models relative to closed models and existing technology, and found the evidence too immature to conclude that open models pose more marginal risk than benefit. It recommended monitoring rather than restriction. Stanford HAI scholars have argued along similar lines, urging policymakers to weigh the benefits of openness while continuing to assess the marginal risks. The debate is real, and thoughtful people land on both sides.
Making the practical call
There is no single winner, only a fit. A short read:
- Choose closed and proprietary when you want the highest capability with the least operational overhead, your data can live with a vendor, and per-token pricing suits your volume.
- Choose open weight when your data must stay on your infrastructure, you want to fine-tune deeply, you want insulation from a vendor’s pricing and lifecycle decisions, and you can run the serving. Verify it is a real permissive license (MIT or Apache 2.0), not a promise or a capped community license.
- Choose fully open source when reproducibility and auditability are the requirement, and you accept a smaller field of qualifying models.
Most organizations end up with a mix, routing different work to different tiers rather than committing to one. The spectrum is the point: the useful question is not "open or closed" but "how open, under what license, for which task."
Frequently Asked Questions
What is the difference between open-source and open-weight AI?
Open weight means you can download the finished model and run it. Open source means you also get the recipe: the training code and enough detail about the data to rebuild and audit the model from scratch. Open weight gives you the product; open source gives you the blueprint. Almost every model marketed as “open source” today is actually only open weight.
Are DeepSeek, Qwen, and Llama open source?
Not in the strict sense. DeepSeek V4 Pro and GLM-5.2 are open weight under the permissive MIT license, which is about as open as licensing gets, but they do not publish full training data and recipe. Qwen’s smaller dense models ship under Apache 2.0, while its Max-tier flagship has historically been closed. Llama is open weight under a custom community license, which the OSI has said is “not open source.”
Which license is truly open for an AI model?
MIT and Apache 2.0 are the two permissive, OSI-recognized licenses that let you use, modify, redistribute, and commercialize a model with no royalty and no user cap. Custom “community” licenses, like Meta’s Llama license, are more restrictive and are not OSI-approved. Before building on any model, read the actual license text rather than the announcement.
Can I run a closed model like GPT-5.6 or Claude on my own servers?
No. Closed, proprietary models such as GPT-5.6, Claude Opus 4.8, and Gemini are available only through the vendor’s API or apps. The weights never leave the vendor’s infrastructure, so self-hosting is not an option. If running the model on your own hardware is a requirement, you need an open-weight model.
Is open-weight AI cheaper than a closed API?
It depends on volume. On hosted-API pricing, open-weight models are often far cheaper per token: GLM-5.2 costs a fraction of Claude Opus 4.8 or GPT-5.6 Sol. But self-hosting swaps per-token pricing for fixed infrastructure cost that only pays off at high, steady volume. At low or bursty volume, a closed hosted API is usually both cheaper and simpler.
Are open models less safe than closed models?
The evidence is genuinely mixed. Open weights raise proliferation concerns, because they can be copied widely and have their safety training removed. But openness also enables independent auditing, red-teaming, and interpretability research that closed APIs block. The 2024 NTIA report framed this as a question of marginal risk and found the evidence too immature to declare open models riskier on balance, recommending monitoring rather than restriction.
Which should a small business choose?
For most small teams, a closed hosted API is the pragmatic default: highest capability, nothing to run, and pay-as-you-go pricing. Reach for an open-weight model when data governance requires keeping information on your own infrastructure, when you need deep customization, or when you want insulation from a vendor’s pricing and deprecation decisions and can support the serving. Many organizations use both, routing each task to the tier that fits.
