Open-Source AI & Open Models Reading List

What it claims

The list carries several positions of its own, stated in Lambert's annotations, that the catalog summary does not mention:

• Size of the gap. The open-closed performance gap is now "roughly 4-6 months." Every leading open model has come from a Chinese lab since about 2024.
• Where open models win. Open models sit on the Pareto cost frontier but not at the absolute performance frontier. The example given is DeepSeek V4 Flash on Artificial Analysis.
• Distillation. This is called "the single most eventful debate around open models in 2026." Lambert's position is that distillation helps Chinese labs but is not the reason they are near the frontier. The panic claiming otherwise "is not grounded in the evidence."
• A revised view of DeepSeek R1. In Apr. 2025 Lambert wrote confidently that DeepSeek did not distill from o1. Lambert now says it is "more possible than I gave it credit" that DeepSeek distilled some o1 traces, because the reasoning-trace extraction methods in Panfilov et al. 2026 make this plausible. Lambert also states that Anthropic confirmed Chinese labs used that technique.
• How frontier labs distill. Lambert says frontier labs have shifted toward logit-based knowledge distillation, first confirmed in Gemini. Open efforts still mostly use supervised fine-tuning on a stronger model's outputs.
• Regulatory risk. One linked piece ("6 months to live for open models," Jul. 2026) argues that vague federal oversight sets up a clash with, or a ban of, frontier open models in the near future.

Evidence

The list makes these points mostly by citing other sources rather than presenting data:

• Gap measurements:
• SemiAnalysis's independent evaluations (Aug. 2026)
• Epoch AI and Artificial Analysis trackers
• Håvard Tveit Ihle's analysis across public and private evaluations (May 2026)
• Release speed. Z.ai's product lead is quoted from 2025: "We open source it within a few hours."
• Misuse. Anthropic's "Detecting and countering misuse of AI: September 2026" report is described as documenting at-scale use of Claude by banned parties. That use mixed SFT-data distillation with routing Claude into products without telling users.
• Regulatory scrutiny of Chinese models. Lawmakers probed these companies, each with a dated news citation:
• DoorDash (Jul. 31 2026)
• Airbnb and Cursor (Apr. 29 2026)
• Apple (May 17 2025)
• Harvey (Aug. 2026)
• Switching to Chinese open models. Thomson Reuters moved off Claude to build on Qwen (Aug. 24 2026). Perplexity adopted DeepSeek R1 (Jan. 28 2025).

How to use it

• Reading order. Each section opens with a core piece, followed by indented supporting pieces. Items marked "[Optional]" are depth, not core:
• the adoption data
• the Pythia/OLMo technical reports
• the 2024 synthetic-data history
• Shortest path to the technical debate: 1. The RLHF textbook chapter on synthetic data and distillation (background) 2. "How much does distillation really matter for Chinese LLMs?" (Feb. 2026) 3. The Panfilov et al. trace-stealing paper 4. "The distillation panic" (May 2026)
• Live data rather than essays. Use the Interconnects Adoption Dashboard (downloads, derivatives and research adoption by region), the Interconnects Artifacts Hub (key models) and The ATOM Report (Apr. 2026).

Limits

• Snapshot date. The list was last updated 15 Sep. 2026, and the author is inviting additions, so contents will shift.
• Single author. Many entries are Lambert's own Interconnects posts, and Lambert's stated purpose is prep for policy-facing writing. The list argues for open models. Most of its safety entries argue that open-weight risk is overstated, with Thinking Machines Lab's piece as the main "take safety seriously" voice.
• Thin cyber section. The author flags the cyber-risk section as underdeveloped ("I plan to develop this further"). It has three entries, all arguing against control or nonproliferation.
• Unchecked figures. The 4-6 month gap figure is the author's synthesis of the linked sources, not a measurement on this page. Check the SemiAnalysis and Epoch sources for the methodology before quoting it.