A reading list dedicated to open models, updated on October 26, 2026, provides an overview of the most useful works for understanding this field of artificial intelligence. Its references cover the reasons for publishing models whose weights are publicly accessible, the gap separating them from closed systems, and distillation, which the list presents as the most significant debate of 2026.
To understand the stakes, the list first cites Mark Zuckerberg’s text, “Open Source AI is the Path Forward,” published in July 2024 regarding the release of Llama 3. It also recommends the study by Irene Solaiman, published in February 2023, which invites viewing models not as open or closed, but on a gradient, depending on licenses, operating costs, or data access. On risks, an article by Sayash Kapoor, Rishi Bommasani and their co-authors, dated February 2024, concludes that text-centric language models only marginally increase documented risks.
On the place of open models in the technological race, the list indicates that the gap with closed models has narrowed in recent years and would now be about four to six months. According to it, the best open models have all come from Chinese labs since around 2024. An analysis by SemiAnalysis, published in August 2026 under the title “Are Open Models Catching Up?”, relies on independent evaluations to reach a similar conclusion. The list also cites a statement from Z.ai’s product lead in 2025 about releasing its models online: “Get it out fast. We open source it within a few hours.”
Western Companies and the Chinese Question
Several US lawmakers have questioned companies about their use of Chinese models. The list mentions DoorDash, in an investigation reported by CNBC on July 31, 2026, as well as Airbnb and Anysphere, the publisher of Cursor, whose case was reported by Bloomberg and Semafor on April 29, 2026. Apple is cited regarding a deal mentioned by the New York Times, reported by Reuters on May 17, 2025, and Harvey appears in August 2026. The list also mentions voluntary switches: Perplexity quickly integrated DeepSeek R1 in January 2025, according to Forbes, and Thomson Reuters relies on Qwen to move away from Claude, according to Business Insider on August 24, 2026.
Distillation at the Heart of the Debate
Distillation involves training a model using the outputs produced by another model. The list dedicates an entire section to it, drawing notably on a report by Anthropic published in September 2026, which describes large-scale use of its products by banned actors, mixing technical distillation and integration of Claude into other services without users being informed. Anthropic confirmed that this technique had been used by Chinese labs.
An article by researchers, including Panfilov and Shumailov, titled “Stealing Reasoning Traces from Proprietary LLM APIs” and dated 2026, further shows that certain programming interfaces of closed labs can allow for the systematic extraction of models’ reasoning traces. The list notes, however, that no clear evidence has been established that DeepSeek distilled OpenAI’s o1 model to develop R1, and it observes that distillation, in its view, detracts nothing from the innovation of the labs concerned.
