Moonshine
On 16 July 2026, Beijing-based Moonshot AI released Kimi K3, a model containing 2.8 trillion parameters and a context window of one million tokens. Its weights have since been published on Hugging Face under a modified MIT licence, making K3 the largest openly available large language model in history. On independent benchmarks from Vals AI it ranks second overall, behind only Anthropic’s Claude Fable 5 and ahead of OpenAI’s GPT-5.6 Sol. On the Arena Frontend Code leaderboard it placed first.
Several headlines have framed this as China closing the AI gap through sheer force of innovation. The reality is more interesting, and considerably less comfortable for everyone involved. In the language of the AI industry, the technique that allegedly accelerated K3’s development is called distillation. In plainer English, there is an older word for an unlicensed spirit produced from someone else’s recipe. The word is moonshine.
In February 2026, Anthropic publicly accused Moonshot AI, DeepSeek and MiniMax of running coordinated distillation campaigns against its Claude model. According to Anthropic, the three labs collectively created approximately 24,000 fraudulent accounts and generated more than 16 million exchanges with Claude, using carefully crafted prompts to extract its most advanced capabilities: agentic reasoning, tool use and coding. These are, by coincidence or otherwise, exactly the capabilities K3 now markets as its defining strengths.
Moonshot was not the only operation under scrutiny. OpenAI sent a memo to the House Select Committee on the Chinese Communist Party on 12 February 2026 accusing DeepSeek of using obfuscated methods and third-party routing to extract capabilities from its models. Google’s Threat Intelligence Group reported disrupting a separate campaign of more than 100,000 prompts aimed at coercing reasoning traces from Gemini. By June, Anthropic escalated further, writing to the Senate Banking Committee to allege that operators connected to Alibaba and its Qwen AI lab had run a second wave involving 25,000 accounts and 28.8 million exchanges.
None of the three labs have issued a substantive public rebuttal. The Chinese Embassy in Washington offered a general denial, stating that Beijing “attaches great importance to the protection of intellectual property rights,” but no company has addressed the specific allegations. The still, it seems, operates more quietly than the revenue officer.
This does not mean K3 is merely a copy. Distillation is a training technique, not a photocopy. Moonshot invested heavily in architecture, scaling and inference optimisation. The model is backed by Alibaba and Tencent, and its engineering team is formidable. But the question of how much of K3’s capability was independently developed and how much was accelerated by systematically harvesting outputs from American models is one the industry has not resolved. The spirit may be real. The provenance is not.
A second narrative being paired with K3 also deserves correction. Several commentators have suggested that Anthropic’s decision to make Fable 5 a permanent part of its premium subscription plans, announced on 17 July, was a competitive response to K3’s launch the day before. The implication is flattering to Moonshot but unsupported. Anthropic’s Fable 5 subscription saga began on 9 June, the day the model launched. It was interrupted by a US government export control order on 12 June that shut down access entirely. When those controls were lifted on 30 June, Anthropic restored access with a temporary window that was extended twice in the face of subscriber backlash, before the company finally made the model permanent on Max and Team Premium plans on 20 July. The timeline was driven by capacity constraints and user pressure, not by anything happening in Beijing.
What is genuinely significant about K3 is not the benchmarks. It is the distribution model. The weights are open. Researchers, startups and sovereign AI programmes from Jakarta to Riyadh can download K3, fine-tune it and deploy it without paying a dollar to an American company or accepting an American terms of service agreement. In a world where frontier AI is increasingly treated as critical infrastructure, that accessibility matters enormously. Moonshine, whatever its origins, flows where the licensed product cannot.
The race for artificial intelligence leadership was never going to be settled by whoever trained the biggest model first. It will be settled by whoever can improve fastest, distribute most widely and convince the world to build on their platform. On that measure, K3 is a legitimate challenge to American dominance, whatever questions remain about how its capabilities were acquired.
But those questions do remain. And any honest account of this moment must include them. A spirit distilled from another maker’s recipe and then given away for free is not the same thing as one brewed from scratch. The distinction matters, not because it diminishes what Moonshot built, but because it reveals the real contest underneath. It is not simply about who has the best model. It is about who owns the knowledge inside it, and whether ownership still counts once the bottle has been opened and poured for the world.