OpenAI’s case for open weights did not begin with gpt-oss. It was already being debated inside the company on October 1, 2022, when Sam Altman sent the board a proposal for a language model that could run on consumer hardware. The email, exposed through the Musk v. Altman lawsuit and published by Simon Willison, is useful because it shows the proposed release as a competitive move, not just a public-interest gesture.
Altman was not proposing to open the most capable model OpenAI had. He described a model with the approximate capability of GPT-3, running locally, and released as open source. The timing was part of the plan: he wrote that OpenAI should do it “before Stability or someone else does.” The stated goals were to discourage other groups from releasing similarly powerful models and make it harder for new efforts to get funded. The email text therefore reads less like a promise to share everything than a proposal to shape the market before somebody else did.
The proposal was tactical, not a release plan
The note opens with Altman saying OpenAI had been having “extensive discussions around open source strategy.” He said the board would discuss it at its next meeting and that the company wanted to act “soon.” Those words matter. They establish an internal proposal and a timetable for discussion, not a record that the board approved the model or that OpenAI shipped it.
The proposed capability level also sets a boundary around what Altman meant by openness in 2022. A GPT-3-class model on consumer hardware would have been a substantial local release, but the proposal was not to publish a frontier model. Its purpose was to get ahead of Stability or another competitor, reduce the incentive to fund rival efforts, and discourage releases that OpenAI considered similarly powerful. All three motives appear in Willison’s transcription and context.
That is a revealing distinction for anyone running models at home today. Open weights can expand access, but a company can also treat them as a strategic lever. The same release can help developers build outside a hosted service while helping the publisher defend its position against other model makers. The 2022 email makes that tension explicit.
OpenAI eventually shipped a different kind of answer
The public admission that OpenAI’s position had changed came on February 1, 2025. In a Reddit AMA quoted by Fortune, Altman said, “I personally think we have been on the wrong side of history here and need to figure out a different open source strategy.” He immediately qualified that statement: “Not everyone at OpenAI shares this view, and it’s also not our current highest priority.” The same article quotes him calling DeepSeek “a very good model” and saying OpenAI would “maintain less of a lead than we did in previous years.”
Six months later, OpenAI released gpt-oss. MIT Technology Review reported that the August 5, 2025 release was OpenAI’s first open-weight language-model release since GPT-2 in 2019. It came in two sizes under the Apache 2.0 license. The smaller model could theoretically run on 16 GB of RAM, while the larger model required a high-end laptop or specialized hardware, according to the same report.
The technical model card makes the release more concrete. OpenAI’s arXiv paper describes gpt-oss-120b and gpt-oss-20b as open-weight reasoning models using an efficient mixture-of-expert transformer architecture. Its abstract says the models were trained using large-scale distillation and reinforcement learning, and that OpenAI released the weights, inference implementations, tool environments, and tokenizers under Apache 2.0.
For local users, the hardware split is the practical point. The gpt-oss-20b model card lists 21 billion parameters with 3.6 billion active parameters and says the model can run within 16 GB of memory. The gpt-oss-120b card lists 117 billion total parameters with 5.1 billion active parameters and says it fits into a single 80 GB GPU such as an NVIDIA H100 or AMD MI300X. Those are publisher-provided deployment targets, not a promise that every consumer computer will run either model comfortably.
Open weights do not remove the safety tradeoff
The move from a 2022 proposal to a released model did not eliminate the hard part of distributing weights. The full model card warns that, once released, determined attackers could fine-tune the models to bypass safety refusals or optimize them for harm, without OpenAI being able to add further mitigations or revoke access. That warning appears in the paper’s limitations, alongside the licensing and deployment details.
That is the tradeoff local-AI readers should keep in view. The 20b card says the smaller model is intended for local or specialized use cases, and its memory target gives users a path to run inference on hardware they control. The open distribution also means control over the weights is no longer limited to the hosted provider. The benefit and the liability arrive together: local deployment can reduce dependence on a remote endpoint, but the publisher has less power to change what released weights can do after the fact.
OpenAI’s model card also says both models were trained on its harmony response format and “should only be used with the harmony format as it will not work correctly otherwise.” That is a reminder that open weights are not the same thing as a frictionless download. The model, inference stack, formatting rules, hardware, and usage policy all shape what a user can actually do with the release. The official model card spells out that dependency rather than hiding it.
What This Means
The 2022 board email shows OpenAI considering open release as a way to move before Stability or another competitor. The 2025 AMA shows Altman later acknowledging that OpenAI had been on the wrong side of the open-source debate. The gpt-oss release turned that changed posture into models that OpenAI describes as open-weight, Apache 2.0 licensed, and usable on local hardware. The same model card warns that releasing weights limits the publisher’s ability to add mitigations later.
For readers choosing between a hosted model and a local one, the important question is not whether “open” sounds good. It is what the release lets you control, what hardware it needs, and what safeguards the publisher can no longer change. The email and the model card point to the same answer from opposite ends: openness is both a user benefit and a strategic decision by the company doing the releasing.
The Bottom Line
OpenAI’s open-weight turn was not a single change of heart. The 2022 email framed a local GPT-3-class model as a defensive move; the 2025 gpt-oss release made local deployment real while preserving serious safety tradeoffs. For local-AI users, the useful question is how much control the weights provide and how much responsibility follows them home.