The biggest models of the summer are being called open source, and mostly they are not. They are open weight: you get the trained weights, not the recipe. For a founder deciding what to build on, the difference is not academic. It decides what you can ship, what you owe, and what can be taken away.
In July 2026, Moonshot's Kimi K3 became the largest open weight model ever released, 2.8 trillion parameters, and topped the Frontend Code Arena ahead of leading closed models, as reported in Tom's Hardware's coverage of the release. Almost every headline called it open source. Then you read the terms.
What you actually get, tier by tier
| Tier | What you get | 2026 example |
|---|---|---|
| True open source | Weights, training code, data recipe, permissive license | ZAYA1 8B, Apertus (Apache 2.0) |
| Open weight, permissive | Weights under Apache 2.0 or MIT, no training recipe | Inkling, GLM 5.2 |
| Open weight, modified terms | Weights under a custom license with conditions | Kimi K3 (Modified MIT), Llama |
| Delayed open | API first, weights promised later | Kimi K3 until July 27 |
| Open tooling, closed model | The agent or IDE is open, the weights are not | Grok Build |
| Open in name only | A research paper and a demo | plenty |
The top tier is rarer than the headlines suggest, and worth knowing by name. The Swiss AI Initiative's Apertus, built by EPFL, ETH Zurich and CSCS, ships 8B and 70B models with open weights, open training data and the full training recipe, the most transparent large release to date and a deliberate answer to European sovereignty concerns, per ExplainX's 2026 Europe Ai landscape review. That is what the word open source actually promises. Almost nothing else on the leaderboard delivers it.
Kimi K3 sits in the middle tiers on purpose. In 2026, the weights ship under a Modified MIT license and were not even downloadable at launch, the release is API only until July 27, per Wan 2.7's licensing breakdown. The training data and pipeline stay private. Compare Thinking Machines' Inkling, released the same week: a 975 billion parameter mixture of experts model with 41 billion active, under plain Apache 2.0, per Simon Willison's write-up. One of these you can build a business on without reading footnotes. The other, read the footnotes.
Why founders keep getting this wrong
Because the word open is doing marketing work. A model company gets community goodwill, benchmark headlines and developer adoption from the word, while a modified license quietly keeps a commercial lever: usage conditions, attribution demands, or clauses that activate at scale. None of that makes the models bad. GLM 5.2 under MIT scored 62.1% on SWE-bench Pro at $1.40 per million input tokens, per devFlokers' June 2026 open source roundup, and that combination of quality and price is genuinely new. The trap is not using these models. The trap is not reading what you agreed to.
The newest trick: open tools, closed models
There is now a subtler variant of the same marketing move: open sourcing the tooling around a model while the model itself stays locked. xAI is the cleanest example. In July 2026 it open sourced Grok Build, its coding agent and terminal interface, per AlphaMatch's coverage of the release, and collected a week of open source goodwill headlines. The model underneath, Grok 4.5, is entirely closed, and it is very good: it ranks fourth on the Artificial Analysis Intelligence Index, above every open weight model, at more than 60 percent below the price of the leading closed models, per Fello AI's launch analysis. Nothing wrong with any of that. But when the word open appears in a headline, check what it is attached to. An open agent driving a closed model gives you none of the things open weights exist to give you: no self hosting, no inspection, no insurance against the provider changing terms.
The category just got its biggest institutional backer. On 24 July 2026, Microsoft published a formal position defining open weight models as "AI models that anyone can download, inspect, modify, and run on their own infrastructure", and argued that competition from open weights "keeps the gains of AI broadly shared rather than concentrated", per Microsoft's open weight brief, a position CEO Satya Nadella has been championing publicly. When a company that sells access to closed frontier models backs the open ecosystem in its own words, the shift stops looking fringe. The license caution still stands: Microsoft is endorsing the category, not any single model's terms.
The regulator just picked a side
If you build from the EU, as we do from Dublin, the open weight versus open source distinction stopped being pedantry this summer: it became a compliance question with fines attached. From 2 August 2026 the European Commission begins enforcing the AI Act's obligations on general purpose model providers, per the Commission's own GPAI guidelines. The Act carves out lighter obligations for models released under a genuinely free and open license, but the bar is specific: the license must allow access, use, modification and redistribution, the weights and architecture must be public, and the provider cannot monetise the model, per Quantamix's breakdown of what is actually exempt.
Then came the ruling that vindicated every license pedant. In January 2026, legal analysis published by the EU AI Office's external advisory group concluded that Meta's Llama Community License does not qualify as free and open under the Act's standard, while Mistral's Apache 2.0 base models do, per Lyceum's 2026 foundation model obligations guide. Read that again: the most downloaded open model family in the world is, in the EU's eyes, not open. The practical takeaway for a founder is simple and slightly boring, which is how good compliance advice always sounds: if the EU is in your market, weights under plain Apache 2.0 or MIT are the safe lane, and every custom community license needs a proper read before it touches production.
The access risk nobody prices in
There is a second, newer risk: the weights you can download today may not be downloadable tomorrow. In July 2026, Reuters reported that China is weighing restrictions on foreign access to its open models, and a Polymarket market prices a US ban on an open source model this year at 16%. The practical hedge is boring: if a model matters to your product, mirror the weights the day you adopt them. Downloaded weights cannot be unreleased.
This is the same discipline we apply when choosing what runs under a client's Ai product: the model is a supplier, and supplier risk is part of the build. Open weight models are the best supplier deal in years. Just read the contract, because open is a spectrum, not a stamp.
Frequently asked questions
Can I use open weight models commercially?
Usually yes, but it depends entirely on the specific license. Apache 2.0 and MIT weights (Inkling, GLM 5.2) are safely commercial. Modified licenses like Kimi K3's need reading, because custom terms can carry conditions plain MIT does not.
What is the difference between open weight and open source?
Open weight means the trained model file is downloadable. Open source, properly used, also includes the training code and data recipe so the work can be reproduced. Most 2026 flagship releases, including Kimi K3, are open weight only.
Does it matter for a small product?
Less at the start, more with success. Scale triggers and redistribution clauses are exactly the terms that stay invisible until your product grows into them. Five minutes with the license file when you adopt beats a forced migration later.
Does the EU AI Act treat open weight models differently?
Yes. From 2 August 2026 the Commission enforces obligations on general purpose model providers, with lighter treatment only for genuinely free and open licenses. The EU AI Office's advisers concluded Llama's community license does not meet that bar, while Apache 2.0 models like Mistral's do.
Is Grok open source?
The models are not. xAI open sourced Grok Build, its coding agent and terminal interface, in July 2026, but the Grok models underneath remain closed and API only. Open tooling around a closed model gives you none of the self hosting or inspection rights that open weights provide.
© 2026 Dinimiciuil Labs. All rights reserved. Written on the build floor in Dublin. You are welcome to quote a short excerpt with a link back; please do not republish the full article without permission.