Hey all! I’ve been prepping for immoderate public-audience and policy-facing penning connected unfastened models, truthful I figured I would stock my investigation materials. There’s tons of awesome worldly successful here.
This is my database of the champion penning connected unfastened models successful the past fewer years. If personification decides they want to get up to velocity connected the area, reference this will beryllium a broad overview of the authorities of affairs. Please remark pieces to see adding below, and I’ll update this complete time.
List past updated: 13 Sep. 2026
Share
What unfastened models are, why group merchandise them, really they subordinate to business strategy, and what the risks are.
On unfastened root AI strategy, a walkthrough of really open-source package has been utilized by businesses and early signs of what that intends for AI — From Open Source Software to Open Source Strategy, Bill Gurley (May 2026).
One of the clearest articulations is Mark Zuckerberg’s comments astir Llama 3’s merchandise arsenic to why Meta releases unfastened models — Open Source AI is the Path Forward, Mark Zuckerberg (Jul. 2024)
Why you should position unfastened models connected a gradient, alternatively than binary open/closed, based connected factors specified arsenic licenses, costs of moving the model, information access, etc. — The Gradient of Generative AI Release: Methods and Considerations, Irene Solaiman (Feb. 2023).
The domiciled unfastened models will play successful the system of the future, arsenic a complement to beardown closed models. Why unfastened models will beryllium utilized to create civilization agentic workflows successful enterprises crossed the world – What comes adjacent pinch unfastened models, Nathan Lambert / Interconnects (Mar. 2026)
A position connected really unfastened models will seizure worth by providing a complementary instrumentality to ample swaths of the existing economy, drafting connected the history of IP and existent debates connected unfastened vs. closed models (e.g. distillation) — Some Simple Economics of Open versus Closed AI, Christian Catalini (Aug. 2026)
Why unfastened models will perpetually beryllium down closed models successful capacity — Open models successful perpetual catch-up, Nathan Lambert / Interconnects (Feb. 2026)
Where take differs for unfastened and closed models — Open and closed models are connected different exponentials, Nathan Lambert / Interconnects (Jun. 2026)
A clear articulation connected really to equilibrium releasing powerful open-weight models while taking information earnestly — A Safe Path to Open Weights, Thinking Machines Lab (Jul. 2026).
Early insubstantial connected marginal risks that showed text-focused LLMs very marginally accrued documented imaginable risks of models — On the Societal Impact of Open Foundation Models, Sayash Kapoor, Rishi Bommasani et al. (Feb. 2024).
Closed models information guardrails are regularly bypassed causing a plethora of existent AI-risks earlier hypothetical risks of unfastened weight models person emerged — The Myth of unsafe Open Source AI, Florian Brand (Jun. 2026).
The wide simplification successful unfastened data, which is simply a important facet that has hampered genuinely unfastened AI investigation — Consent successful Crisis: The Rapid Decline of the AI Data Commons, Shayne Longpre et al. (Jul. 2024).
Recent examples connected really beardown Chinese models effect the AI ecosystem — Kimi K3: The open-weights escalation, Nathan Lambert / Interconnects (Jul. 2026) / GLM-5.2 is the measurement alteration for unfastened agents, Nathan Lambert / Interconnects (Jun. 2026).
A summary of the communicative of unfastened models successful 2025: Nathan Lambert connected China’s AI Ecosystem and the Open Model Gap | The Curve 2025, Golden Gate Institute for AI (Nov. 2025).
[Optional] Latest information connected unfastened exemplary adoption:
A wide summary connected US vs. China exemplary take — The ATOM Report (Apr. 2026),
The latest information connected exemplary downloads, derivatives, and investigation take by region — Interconnects Adoption Dashboard, and
The astir important models to cognize astir successful the ecosystem — Interconnects Artifacts Hub
Who is starring successful unfastened models, really this has changed complete time, really China maintains its starring position, and applicable history.
Why the U.S. needs to put successful unfastened models for basal R&D / invention successful the look of increasing title from China – The ATOM Project, Nathan Lambert (Aug. 2025)
The lens arsenic to why unfastened models thief spur investigation invention and beneficial outcomes for AI — Why I build unfastened connection models, Nathan Lambert / Interconnects (Oct. 2024)
Why unfastened models foster education, invention and competition, 3 halfway American values — Banning Open Source AI Would Be A Mistake, Nathan Lambert & Kevin Xu (Jun. 2026)
Why the caller “vibe regulation” / vague national oversight mechanisms group america up for a conflict and-or prohibition of frontier unfastened models successful the adjacent early — 6 months to unrecorded for unfastened models, Nathan Lambert / Interconnects (Jul. 2026)
[Optional] Fully unfastened connection exemplary method reports to exemplify the commencement of the creation successful understanding: Pythia (EleutherAI, 2023), Olmo (2024), Olmo 2 (2024), Olmo 3 (2025)
Chinese open-source history starring up to AI — Chinese Open Source: A Definitive History, Kevin Xu (Mar. 2026).
China’s structural advantages successful open-source — China’s Structural Advantage successful Open Source AI, Kevin Xu (Jun. 2025).
How Chinese labs themselves talk building models, and really the Chinese manufacture differs from the U.S. — Notes from wrong China’s AI labs, Nathan Lambert / Interconnects (May 2026).
Why Chinese labs are truthful bully astatine keeping up pinch American title (e.g. American unfastened weight labs struggle to compete pinch Chinese labs connected adjacent capacity comparisons) — GLM-5.3: How Chinese labs support stride pinch the frontier, Nathan Lambert / Interconnects (Aug. 2026).
Prominent uses of Chinese models by Western companies person prompted meaningful regulatory attraction (more discussion)
Lawmakers person probed the pursuing companies complete utilizing Chinese models: DoorDash (CNBC, Jul. 31 2026), Airbnb (Bloomberg, Apr. 29 2026; Semafor, Apr. 29 2026), Anysphere / Cursor (Bloomberg, Apr. 29 2026; Semafor, Apr. 29 2026), Apple (Reuters, May 17 2025)
Other occidental companies person very publically shifted the models they usage from American, closed labs to Chinese unfastened models to prevention costs. Examples see Perplexity prominently and quickly adopted DeepSeek R1 (Forbes, Jan. 28 2025) and Thomson Reuters building connected Qwen to move disconnected Claude (Business Insider, Aug. 24 2026)
Leave a comment
What is distillation and really overmuch does it thief Chinese labs, really do unfastened models effect frontier AI risks for illustration cybersecurity, and really acold are unfastened models down the closed frontier?
The open-closed exemplary spread has reduced successful caller years, and is now astatine astir 4-6 months. The starring unfastened models person each travel from Chinese labs since ~2024.
SemiAnalysis article which ran independent evaluations, concluding that unfastened models person been getting person to the person frontier of capacity complete clip — Are Open Models Catching Up?, SemiAnalysis (Aug. 2026)
Open models are connected the Pareto cost frontier, while not astatine the absolute capacity frontier. E.g. DeepSeek V4 Flash, spot evaluation and costs connected Artificial Analysis.
Data sources from Epoch AI and Artificial Analysis (and U.S. v China, related) showing the open-closed spread complete time.
An independent study of the open-closed spread crossed a operation of nationalist and backstage evaluations — How acold down are unfastened models?, Håvard Tveit Ihle (May 2026)
E.g. successful 2025, the merchandise lead of Z.ai said pinch respect to their merchandise clip “Get it retired fast. We unfastened root it wrong a fewer hours.” — The Z.ai Playbook, ChinaTalk (Nov. 21, 2025)
Cyber, risks & unfastened models (I scheme to create this further)
Why we cannot efficaciously prohibition unfastened models arsenic utilized by bad actors for cyber capabilities (they will ever person access) — The OpenAI/Huggingface incident; really we should negociate the imminent presence of autonomous hacking excessively inexpensive to meter, Joshua Saxe (Jul. 2026)
What the authorities should do to observe, orient, decide, and enactment pinch respect to emerging cyber threats (versus blocking models based connected in-house capacity assessments) — We urgently request a coherent nationalist AI cybersecurity policy, Joshua Saxe (Aug. 2026)
Why you cannot expect to power entree to AI astatine a definite period (e.g. unfastened weight models) and request to hole nine to tackle risks downstream of disposable intelligence — Nonproliferation is the incorrect attack to AI misuse, Helen Toner (Apr. 2025)
Distillation – the process of training connected output tokens from different exemplary – is the azygous astir eventful statement astir unfastened models successful 2026.
For basal background, spot a textbook chapter connected synthetic information & distillation generally, from Reinforcement Learning from Human Feedback (post-training textbook published successful 2026)
How distillation helps the Chinese labs, but doesn’t return distant from their invention — How overmuch does distillation really matter for Chinese LLMs?, Nathan Lambert / Interconnects (Feb. 2026)
A very transparent archiving of really Chinese institution usage Anthropic’s products and circumvent the position of work aliases intended use. The study specifications at-scale usage of Anthropic’s products by banned parties, arsenic a operation of method distillation (mentioned via SFT data) and extended routing of Claude into their products and services without telling users —Detecting and countering misuse of AI: September 2026.
A caller insubstantial that showed that the frontier labs had implementations successful their APIs that made systematic extraction of reasoning traces (the important portion of modern training) done clever tricks. Recent distillation paper, my penning connected it — Stealing Reasoning Traces from Proprietary LLM APIs, Panfilov, Schmotz, Shumailov et. al 2026 (more connected X). Anthropic confirmed this method was utilized by Chinese labs.
Why the governmental panic complete distillation, claiming that distillation is the only logic Chinese models are adjacent to the frontier, is not grounded successful the grounds — The distillation panic, Nathan Lambert / Interconnects (May 2026)
How labs tin usage distillation to amended models successful an era of scaling RL environments crossed agentic behaviors — How distillation is utilized coming and what capacity uplift it gives to unfastened models, Nathan Lambert (Jul. 2026)
[Optional] More history: In 2024, I wrote Frontiers successful synthetic data wherever the cardinal points were that synthetic data, chiefly successful “distilling” models by training pinch SFT connected outputs from a stronger model, was the ascendant shape of distillation. Frontier labs had been shifting the logit-based, knowledge distillation, confirmed earliest successful Gemini and continuing to this day. In early 2025, location was important statement connected if DeepSeek-R1 was distilled from OpenAI’s o1 model. There is nary clear grounds suggesting that they did, and successful Apr. of 2025 I wrote confidently that DeepSeek did not distill. At the clip of R1, it is much imaginable than I gave it in installments to that DeepSeek did distill immoderate o1 traces to make it easier for them to train their R1 exemplary – based connected the supra reasoning trace extraction methods. This does not return distant from the invention of it, but it’s worthy being realistic and is simply a measurement that distillation could accelerate China closing the spread to American labs.
English (US) · 
Indonesian (ID) · 
