Open-weight vs open-source AI models: what the difference is, which models qualify, and why IBM's CEO says open weights now win on cost

Open-weight models vs open-source models is a distinction most AI announcements blur. An open-weight model publishes its trained parameters, so you can download it, run it on your own hardware and fine-tune it. An open-source AI model, as the Open Source Initiative defines it, goes further: it also releases the code used to train it and enough information about its training data that a skilled person could build a substantially equivalent system. By that standard, nearly every popular "open" model released in 2026 is open-weight but not open source. The question got fresh attention on October 21, 2026, when IBM CEO Arvind Krishna told analysts that open models had become cheap and good enough for most enterprise work. Licenses are a separate layer, and our open-weight license comparison covers them in detail.

This explainer is based on the OSI's Open Source AI Definition 1.0 and its FAQ, the Hugging Face model cards and license files for Llama 4, Qwen3.8, Mistral Small 4, Gemma 4, DeepSeek V4.1 Flash, Reflection Beam and Kimi K3, IBM's third-quarter results filing, a transcript of IBM's October 21, 2026 earnings call, and The Register's October 22 report, checked October 27, 2026.

The short version

• Open weights means the trained parameters are public. You can run and fine-tune the model, but you may not be able to see how it was made or rebuild it.
• Open-source AI, under the OSI definition, means the weights, the full training and inference code, and detailed data information, all under terms that let anyone use, study, modify and share the system.
• Releasing the training data itself isn't required by the OSI definition, but describing it in detail is.
• Most big releases fall short. The model cards for Gemma 4, Qwen3.8-27B, Mistral Small 4, DeepSeek V4.1 Flash and Reflection Beam use open-source licenses but don't release training code or datasets. Llama 4's and Kimi K3's licenses add their own conditions.
• IBM's pitch: Krishna said open models cost 25% to 10% of frontier token prices and are "90 percent as effective."

What's the difference between open-weight and open-source AI models?

The difference is what you get beyond the model file. A model's weights are the billions of numbers learned in training; with them and some inference code, you can run the model. That's what "open weights" promises, and it's what most labs mean when they call a model open.

Open source, in software, means something more specific: access to the source code, the form a developer would actually use to change the program, under a license that lets anyone use, study, modify and share it. For AI, weights alone work much like a compiled program. You can run and adjust them, but you can't see what went into them, check them for problems in the data, or rebuild them differently.

Two separate questions follow, and they're easy to mix up:

• Is the license open? Apache 2.0 and MIT are OSI-approved licenses. Meta's Llama license and Moonshot's Kimi license add conditions on who can use the model and how, as our license comparison details.
• Is the release complete? Even under Apache 2.0, a release that ships weights without training code and data information isn't open-source AI under the OSI definition.

What does the Open Source AI Definition require?

The OSI published version 1.0 of its Open Source AI Definition on October 28, 2024. It requires that an AI system be made available under terms granting four freedoms: to "use the system for any purpose and without having to ask for permission," to "study how the system works and inspect its components," to "modify the system for any purpose, including to change its output," and to "share the system for others to use with or without modifications, for any purpose."

To make those freedoms practical, the definition requires the "preferred form" for making changes, which for machine-learning systems has three parts:

• Data information: "sufficiently detailed information about the data used to train the system," enough for a skilled person to build a substantially equivalent system. That includes listing publicly available training data and where to get it, and third-party data and where to obtain it, including for a fee.
• Code: "the complete source code used to train and run the system," covering data processing and filtering, training, validation, inference and model architecture, under OSI-approved licenses.
• Parameters: "the model parameters, such as weights or other configuration settings," under OSI-approved terms.

The training data itself doesn't have to be released. The OSI's FAQ says "training data does not equate to a software source code," and that data which can't legally be shared, such as some personal or licensed data, can be described instead.

Which AI models are actually open source?

Few of the models people call open meet the OSI bar. The OSI's FAQ lists the models its volunteers found to pass when they tested the definition: EleutherAI's Pythia, Ai2's OLMo, LLM360's Amber and CrystalCoder, and Google's T5. Those that didn't pass included Meta's Llama 2, Grok, Microsoft's Phi-2 and Mistral's Mixtral, for missing components, incompatible legal terms or both. The OSI says these results "are not certifications of any kind."

For today's widely used models, the Hugging Face cards tell the story.

Last verified: October 27, 2026

The license conditions differ too. Llama 4's download page on Hugging Face is gated: you must accept Meta's terms and give your full legal name, date of birth and organization. Kimi K3's license requires model-as-a-service companies whose aggregate revenue exceeds $20 million over any 12 consecutive months to sign a separate agreement with Moonshot, and requires products with more than 100 million monthly active users or more than $20 million in monthly revenue to display "Kimi K3" prominently, though internal use is exempt. Moonshot's newer Kimi K3.1 Pro has no public weights yet; our Kimi K3.1 Pro story tracks the promised release. Reflection's Apache 2.0 release of Beam is covered in our Beam story.

None of this makes these models less useful to run. It means you can use and adapt them but can't fully inspect or reproduce how they were built, and that "open source" in a press release usually means open weights.

Why does the difference matter?

It matters for three practical reasons.

Auditing. Without data information and training code, you can't check what a model learned from, whether benchmark answers leaked into its training data, or why it behaves as it does. You can only test its outputs.

Legal terms. A custom license can limit who uses the model, require branding or bar some uses. An OSI-approved license like Apache 2.0 or MIT doesn't. That decides whether a large company, a cloud reseller or a startup can ship a product on the model without a separate deal.

Regulation. The EU AI Act treats models released under free and open-source licenses differently for some obligations, and how a license is written can decide whether that applies. Our license comparison covers the European Commission's guidance.

For running a model yourself, the weights and a permissive license are what count. Our Ollama guide shows how to run open-weight models locally.

What did IBM's CEO say about open-weight models?

On IBM's third-quarter earnings call on October 21, 2026, Krishna answered an analyst's question about how enterprises weigh cost against control. According to The Register's October 22 report by Simon Sharwood and a published transcript of the call, he said:

• On the gap: on a broad benchmark in late 2024, "the top model might score a 70 and the open model might score a 20 or a 30," but in late 2025 "maybe the gap was not 70 to 20 or 30, it was 70 to 50."
• On cost: open models run at "anywhere from 25 percent to 10 percent of the token cost of the frontier models," which is "very attractive given that they're now 90 percent as effective."
• On the choice: "Would you like something that's 10 percent of the cost if it's 90 percent as good as the best, or do you want to pay that 10 times higher cost?" Most tasks, he said, such as "summarizing an email, examining a contract, answering a customer call," can be done "with the much cheaper answer."
• On the catch: running them well needs specialists. "Where are my kernel engineers? Where are the people who worry about the underlying network?"
• On scale: he estimated a large enterprise could start on-premises with "two or three racks."

Krishna didn't name the benchmark behind his 70, 50 and "90 percent" figures, so they're best read as his characterization rather than a measured result. IBM also has a stake in the answer: Krishna said on the call that IBM's mainframe business would benefit from on-premises AI demand "with our Spyre cards as time goes on," and The Register reported that he expects it to lift Spyre sales. IBM reported third-quarter revenue of $17.1 billion, up 5%, in its October 21 filing.

Whether the cost claim holds depends on your volume and setup. Our cost comparison of frontier APIs and open models works through when self-hosting pays off, and our analysis of 2026's API price cuts shows the frontier labs cutting their own prices at the same time.

What's confirmed, reported and uncertain

• Confirmed: the OSI definition's text, date and FAQ; the license listed on each model card and what each card says about training code and data; the Kimi K3 license terms; Llama 4's gated access; IBM's third-quarter revenue.
• Reported: Krishna's quotes, from The Register's report and a third-party transcript of the call (IBM's own recording wasn't reviewed); his Spyre comment.
• Uncertain: the benchmark behind Krishna's "90 percent" figure; whether any 2026 frontier-scale model meets the OSI definition (the OSI hasn't published a newer validation list); the license Kimi K3.1 Pro will use.