Open-Source AI and Open Models: A Reading List

List last updated: 11 September 2026

This reading list brings together writing on open models from the past several years. It is organized as an overview of the field, covering the foundations of open-model strategy, US-China competition, technical progress, safety, cybersecurity, and distillation.

Foundations

These selections explain what open models are, why organizations release them, how they relate to business strategy, and which risks they create.

• Open-source AI strategy: Bill Gurley’s From Open Source Software to Open Source Strategy (May 2026) examines how businesses have used open-source software and what early developments suggest for AI.
• Why Meta releases open models: Mark Zuckerberg’s Open Source AI is the Path Forward (July 2024) presents Meta’s rationale for releasing open models around the launch of Llama 3.
• A spectrum of openness: Irene Solaiman’s The Gradient of Generative AI Release: Methods and Considerations (February 2023) argues that generative AI releases should be assessed along a gradient, considering factors such as licenses, operating costs, and access to data.
• The future role of open models: Nathan Lambert’s What comes next with open models (March 2026) considers open models as a complement to powerful closed models and discusses their potential use in customized enterprise agent workflows.
  • Open models in perpetual catch-up (February 2026) examines why open models may remain behind closed models in performance.
  • Open and closed models are on different exponentials (June 2026) discusses differences in adoption between open and closed models.
• Safety and open weights: Thinking Machines Lab’s A Safe Path to Open Weights (July 2026) addresses how powerful open-weight models might be released while maintaining attention to safety.
  • Sayash Kapoor, Rishi Bommasani, and colleagues’ On the Societal Impact of Open Foundation Models (February 2024) presents an early analysis of marginal risks, finding that text-focused large language models produced only marginal increases in documented potential risks.
• Declining access to open data: Shayne Longpre and colleagues’ Consent in Crisis: The Rapid Decline of the AI Data Commons (July 2024) describes the reduction in openly available data and its effect on open AI research.
• Chinese models and the ecosystem: Nathan Lambert’s Kimi K3: The open-weights escalation (July 2026) and GLM-5.2 is the step change for open agents (June 2026) examine how strong Chinese models are affecting the broader AI ecosystem.
  • Nathan Lambert on China’s AI Ecosystem and the Open Model Gap | The Curve 2025, published by the Golden Gate Institute for AI (November 2025), summarizes the development of open models in 2025.

Optional data sources

• The ATOM Report (April 2026) provides a general summary of model adoption in the United States and China.
• The Interconnects Adoption Dashboard tracks model downloads, derivatives, and research adoption by region.
• The Interconnects Artifacts Hub catalogs significant models in the ecosystem.

US-China Competition

This section covers leadership in open models, changes over time, China’s position, and the historical context behind current developments.

• The case for US investment: Nathan Lambert’s The ATOM Project (August 2025) argues that the United States should invest in open models for fundamental research and innovation amid growing competition from China.
  • Why I build open language models (October 2024) explains how open models can support research innovation and other beneficial outcomes.
  • Banning Open Source AI Would Be A Mistake, by Nathan Lambert and Kevin Xu (June 2026), discusses open models in relation to education, innovation, and competition.
  • 6 months to live for open models (July 2026) considers how vague federal oversight mechanisms could lead to conflict over, or restrictions on, frontier open models.
  • Optional technical reports: The reports for Pythia (EleutherAI, 2023), Olmo (2024), Olmo 2 (2024), and Olmo 3 (2025) provide examples of fully open language-model development.
• China’s open-source history: Kevin Xu’s Chinese Open Source: A Definitive History (March 2026) traces Chinese open-source development before the rise of modern AI.
  • China’s Structural Advantage in Open Source AI (June 2025) examines China’s structural advantages in open-source AI.
  • Nathan Lambert’s Notes from inside China’s AI labs (May 2026) describes how Chinese labs approach model development and how the industry differs from that of the United States.
  • GLM-5.3: How Chinese labs keep stride with the frontier (August 2026) examines how Chinese labs continue to compete with the frontier, including comparisons involving American open-weight labs.
• Regulatory attention to Chinese models: Use of Chinese models by Western companies has prompted regulatory scrutiny.
  • Lawmakers questioned DoorDash (CNBC, July 31, 2026), Airbnb (Bloomberg, April 29, 2026, and Semafor, April 29, 2026), Anysphere, the company behind Cursor (Bloomberg, and Semafor), and Apple (Reuters, May 17, 2025).
  • Some Western companies have shifted from American closed models to Chinese open models for cost reasons. Perplexity adopted DeepSeek R1 (Forbes, January 28, 2025), while Thomson Reuters built on Qwen as part of a move away from Claude (Business Insider, August 24, 2026).

Technical Details

This section focuses on the performance gap between open and closed models, distillation, and the cybersecurity implications of open models.

The open-closed performance gap

The gap between leading open and closed models has narrowed in recent years and is estimated at roughly four to six months. Since approximately 2024, the leading open models have come from Chinese labs.

• Are Open Models Catching Up?, by SemiAnalysis (August 2026), reports independent evaluations indicating that open models have moved closer to the closed frontier over time.
• Data from Epoch AI and Artificial Analysis tracks the gap over time. Artificial Analysis also provides related US-China comparisons.
• Håvard Tveit Ihle’s How far behind are open models? (May 2026) analyzes the gap using a mixture of public and private evaluations.
• A 2025 statement from the product lead of Z.ai, quoted in The Z.ai Playbook by ChinaTalk (November 21, 2025), described a rapid release approach: “Get it out fast. We open source it within a few hours.”

Cybersecurity, risks, and open models

• Joshua Saxe’s The OpenAI/Huggingface incident; how we should manage the imminent arrival of autonomous hacking too cheap to meter (July 2026) argues that open models cannot effectively be banned from use by malicious actors because those actors will retain access.
• We urgently need a coherent national AI cybersecurity policy (August 2026) discusses how governments could observe, assess, decide, and act on emerging cyber threats rather than relying on internal capability assessments to block models.

Distillation

Distillation, the process of training one model on output tokens generated by another, became one of the most prominent debates surrounding open models in 2026.

For introductory material, the synthetic data and distillation chapter in Reinforcement Learning from Human Feedback, a post-training textbook published in 2026, provides background on the subject.

• Nathan Lambert’s How much does distillation really matter for Chinese LLMs? (February 2026) examines how distillation benefits Chinese labs without eliminating the role of their own innovation.
• The paper Stealing Reasoning Traces from Proprietary LLM APIs, by Panfilov, Schmotz, Shumailov, and colleagues (2026), reports that frontier labs had API implementations that allowed systematic extraction of reasoning traces through carefully designed techniques. Anthropic later stated that Chinese labs had used this technique in its September 2026 threat intelligence report.
• Nathan Lambert’s The distillation panic (May 2026) argues that claims portraying distillation as the sole reason Chinese models approached the frontier are not supported by the available evidence.
• How distillation is used today and what performance uplift it gives to open models (July 2026) discusses how labs use distillation as reinforcement-learning environments scale across agent behaviors.

Additional history

In Frontiers in synthetic data, published in 2024, Nathan Lambert described synthetic data, particularly supervised fine-tuning on outputs from stronger models, as the dominant form of distillation. Frontier labs had also been moving toward logit-based knowledge distillation, a direction first confirmed in Gemini and continued afterward.

In early 2025, researchers debated whether DeepSeek-R1 had been distilled from OpenAI’s o1 model. There was no clear evidence that this had occurred, and an April 2025 analysis argued that DeepSeek had not distilled o1. Later developments involving reasoning-trace extraction made it more plausible that DeepSeek might have used some o1 traces to simplify training R1. That possibility does not remove the model’s innovation, but it illustrates one way distillation could help narrow the gap between Chinese and American labs.