Open-Source AI & Open Models Reading List

Nathan Lambert published a reading list titled "Open-Source AI & Open Models Reading List," describing it as a compilation of the best writing on open models from recent years. The list was last updated 15 Sep. 2026, and Lambert said he prepared it while working on public-audience and policy-facing writing on open models, inviting readers to comment with pieces to consider adding.

The list is organized into sections. "Foundation" covers what open models are, why organizations release them, business strategy and risks, citing works including Bill Gurley's "From Open Source Software to Open Source Strategy" (May 2026), Mark Zuckerberg's "Open Source AI is the Path Forward" (Jul. 2024) on Meta's reasons for releasing open models, and Irene Solaiman's "The Gradient of Generative AI Release" (Feb. 2023).

"US-China Competition" addresses who leads in open models, how China maintains its position and relevant history. Lambert states that leading open models have all come from Chinese labs since roughly 2024. He also lists Western companies questioned by lawmakers over using Chinese models: DoorDash (CNBC, Jul. 31, 2026), Airbnb and Anysphere/Cursor (Bloomberg and Semafor, Apr. 29, 2026), Apple (Reuters, May 17, 2025) and Harvey (Aug. 2026). He cites Perplexity's adoption of DeepSeek R1 (Forbes, Jan. 28, 2025) and Thomson Reuters building on Qwen to move off Claude (Business Insider, Aug. 24, 2026).

"Technical Details" covers distillation, cybersecurity and the open-closed performance gap, which Lambert says has narrowed to roughly 4-6 months. He cites SemiAnalysis's "Are Open Models Catching Up?" (Aug. 2026) and data from Epoch AI and Artificial Analysis. On distillation, he references "Detecting and countering misuse of AI: September 2026" and a paper on stealing reasoning traces, noting Anthropic confirmed the technique was used by Chinese labs. The list also includes an optional history section on synthetic data and the early-2025 debate over whether DeepSeek-R1 was distilled from OpenAI's o1, where Lambert writes there is no clear evidence that it was.