Powerful Language Models Released for Public Use
The most notable large-scale language models that are currently available for public download include: Evaluation of Artificial Analytics Intelligence Index, GLM-5.2, MiniMax-M3, DeepSeek V4 Pro, and Kim K2.6. Intriguingly, all these models originate from China. They have emerged from well-funded labs and are distributed under permissive licenses, suggesting an open-access framework. However, while users can download the model weights, run them locally, and modify them to develop new products, the inner workings—such as training data and coding techniques—remain largely inaccessible. It’s a bit like having an open door, but peering inside reveals very little.
Definitions surrounding open-source software are strict. Per the OSI’s criteria, open-source AI systems are expected to provide ample data, code, and parameters to allow users to study, modify, and share the underlying systems. On the other hand, “open weights” merely reveal the results of the training process—the numerical parameters—akin to a cookbook offering only a list of ingredients in ready-made meals. You can enjoy the food, but the chance to prepare it for yourself is limited.
For organizations looking to implement models for something like enterprise search or code generation, these “frozen dinners” are not just practical conveniences; they need to be cost-effective and accessible without locking into a subscription model. For instance, DeepSeek V4 Pro charges only 4 cents per task, which is substantially cheaper—over 20 times lower than GPT 5.5 and more than 40 times less than Claude Opus 4. At such rates, concerns about scrutinizing training data may seem rather trivial.
In terms of engineering, these models are quite impressive. GLM-5.2 operates with a staggering 744 billion total parameters and 40 billion active parameters for each token. Its architecture employs a method called Expert Mix, which allows for extensive capability without being resource-intensive. DeepSeek V4 Pro, in particular, boasts a total of 1.6 trillion parameters, effortlessly handling up to 1 million context tokens. That’s like having an entire codebase or a sizeable library accessible in real-time. Users have the flexibility to opt for different levels of cognitive processing—think a little, or a lot, or not at all. The machine adapts to your preferences.
When It’s Not Free
There’s an interesting parallel to the notion of open models in the realm of free software. Eric Raymond famously discussed this in “The Cathedral and the Bazaar.” Yochai Benkler explored commons-based peer production, while Christopher Kelty analyzed free software communities as collaborative systems. These frameworks illustrate how an ecosystem of open models can flourish, where enthusiasts refine and reveal the worth of exposed weights for reputation and personal satisfaction.
Yet, this analogy falters in crucial aspects. In the past, the number of volunteers in the bazaar overwhelmingly surpassed those in the cathedral. In contrast, by 2026, the open model ecosystem appears designed for the cathedral to release its products through the bazaar. For example, DeepSeek recently completed a staggering funding round exceeding $7 billion, while Moonshot AI has raised about $2 billion. Companies like Alibaba are integrating these models into commerce and robotics, continuously investing in their Qwen product line. These are not merely volunteer efforts; they’re established industrial entities focused on a strategy of “strategic openness” as a tool for ecological capture, price disruption, and geopolitical positioning. The U.S.-China Economic Security Review Commission has remarked that China is committed to an open model strategy to spur adoption and enhance iterations through open publishing and aggressive pricing.
A significant paradox arises in this field: openness can enhance participation at the edges but may simultaneously centralize power. Observations by Aaron Shaw and Benjamin Mako Hill indicate that as collaborative systems grow, they often trend toward oligarchy. The open model reflects similar patterns in the industry—while anyone can download weights, very few can produce them. As the open web becomes more inclusive, the gap between utilizing a model and fully comprehending it will likely continue to expand.
The Cookbook’s Secrets
Contrastingly, older open projects maintain a lower profile. Take Ai2’s OLMo program, for instance. It provides complete access to training data, training code, and reproducible recipes, which puts it at the top of openness indexes, although it may rank lower in feature benchmarks. This isn’t by chance; true transparency often entails substantial costs and a willingness to be audited, replicated, and exposed to correction. Institutions seeking to dominate benchmarks often lack the inclination for such openness.
Thus, the term “open” now describes two diverging projects. One focuses on functional access—enabling users to operate powerful models without relying on paid APIs. The other emphasizes knowledge access—the right to understand a model’s origins, training materials, and operational rationale. These definitions coexisted comfortably when the leading open models were also the most transparent. That balance no longer holds. Ready-made meals can be appealing, but the recipe is still under wraps.
Looking ahead, the next few years will likely revolve around open weights for infrastructure, particularly in coding, agency work, and enterprise contexts. The competition over what “open” truly means is just beginning.
