In a newsletter dedicated to open models, a researcher who contributed to the creation of the Olmo models at AI2 compares the current ecosystem to major open-source software, like the Linux operating system. According to him, this comparison does not fully capture the dynamics of these models, whose economic future remains uncertain.
The most frequently cited analogy faces a distinction. A strictly open language model, meaning one delivered with its complete training recipe, its data, and its code, would be closer to a free operating system. Open-weight models, which provide only the learned parameters and the code to run them, more closely resemble specific versions of software installed in a project.
A model’s weights have an average short lifespan, but they remain in use for a long time, like many widely used software packages. This is why, according to the author, many companies continue to run their workflows on Llama 3, even though agentic behaviors, where a model autonomously chains actions, developed several years later. The open recipe, illustrated today by Olmo and its predecessors, like EleutherAI’s Pythia, is a resource-intensive process that any company can take up, modify, and relaunch to produce a new set of weights. Improvements made by the community on the data or training code can, in the best cases, fuel the next generation.
It is within this framework that Nvidia is heavily investing in nearly open models. For its Nemotron models, the company publishes all data it can legally distribute, as well as the training code. The author sees in this a desire for a world where many people could build their own “token machines,” so that intelligence is not monopolized—a world that assumes strong demand for inference from many companies. A token is a unit of text processed by a model, and inference refers to using the model to produce responses.
A Demanding Economy
Developing competitive models remains extremely costly, but access to this race has remained broader in the industry than many would have imagined, according to the author. He notes that most observers consider training too expensive and the open recipe too far behind for a new lab focused on a part of LLM training to be viable. However, the withdrawals of Databricks and 01.ai seem to him isolated cases rather than a sign of a general trend.
Two trajectories are emerging. In the first, the approach works: according to Wired, Nvidia would commit $26 billion to this venture, and if this bet succeeds, the demand for its chips and its profits would exceed the cost of building the models. The author believes the open ecosystem will increasingly depend on Nvidia’s funding in the coming years, and that the economic remuneration of these models will ultimately have to be proportional to the revenue generated by Anthropic’s and OpenAI’s APIs.
Two Possible Trajectories
In the second, if none of these financial paths materialize, open models would follow a different path from the most advanced closed models, focused on efficiency, modifiability, and specialization. The author considers this scenario the most likely: open models would remain very useful but would mainly occupy a “long tail” compared to closed models, which hold dominant positions in areas like knowledge work, drug discovery, or software development. This long tail could take the form of specialized agents, deployed on-premises, on private data and repetitive tasks.
The author also emphasizes that training is becoming more complex and abstract. The current rise of open models largely relies on post-training, meaning the adaptation of models like DeepSeek V4 Flash, Inkling Small, or GLM 5.X to specific agentic tasks, often via fine-tuning. Tinker, which he says is the most popular fine-tuning interface today, illustrates this practice. But the ability to train a base model into a general agentic reasoner is becoming, in his view, as opaque as large-scale pre-training practices were a few years ago. The usual vocabulary, based on pre-training, mid-training, and post-training, could evolve into pre-training, reasoning training, and post-training.
The number of actors publishing base models is decreasing, and some publishers are experimenting with revenue-sharing licenses for using their models in products or for inference, like those of Kimi-K3 and Qwen3.8-2.4T-A95B. Finally, the author mentions the case of Meta, which could release its weights under an open license: in his opinion, the release of Muse Spark 1.2, a highly performant model, would curb the revenue growth of Anthropic and OpenAI, which rely on selling tokens. Nvidia wants to teach everyone to fish for tokens, while Meta floods the market with tokens.
