Sam Altman: OpenAI plans to release a local GPT-3-level model soon

Bulletin: On Local Minds

The Specola’s log records a dispatch of note. According to a post by Simon Willison dated July 20, 2026, an internal OpenAI communication attributed to Sam Altman surfaced via the account techemails, quoting internal deliberations on open source strategy.

The quoted passage states that OpenAI has been holding extensive discussions on this question and intends to raise it further at the organization’s next board meeting. It further quotes a stated aim: to build a language model with the approximate capability of GPT-3, sized to run locally on consumer hardware, and to release it. The quoted text specifies a motive of timing, that OpenAI would prefer to do this soon, before Stability or another party does so first. The excerpt closes on a partial line suggesting a broader rationale, that releasing such a model would help discourage others from releasing similar systems themselves, though the quoted material available to this desk cuts off before the sentence completes.

Willison’s post frames the excerpt as a notable historical artifact of OpenAI’s internal thinking on open weights. No further context, date of the original communication, or confirmation from OpenAI accompanies the quoted material as summarized here.