At 00:00 UTC today, Moonshot AI published the full weights for Kimi K3, and the open-weight tier stopped being a discount option.

K3 is 2.8 trillion parameters, the largest open-weight release in history, and it now sits on Hugging Face under Moonshot's own Kimi K3 License for anyone to download, self-host, and modify. Read that license before you build on it: it is MIT-like, but it is a bespoke document rather than the Modified MIT everyone assumed would carry over, and reporting on the terms describes commercial gates for anyone hosting K3 as a paid service above a revenue threshold. Eleven days ago it was an API you rented. Today it is a file you own.

That is the entire conversation on X right now, and for good reason. But there is a number circulating alongside it that deserves better handling than it is getting, because the panic version and the honest version are very different. Here is what actually shipped, and what the benchmarks do and do not tell you.

Start with what makes this different from every previous open-weight release. K3 is not a small model chasing a big one. On the Artificial Analysis Intelligence Index it scores 57.1, against Fable 5 at 59.9 and GPT-5.6 Sol at 58.9. That is a gap measured in a couple of points, not a generation, and it is the closest an openly downloadable model has ever come to the closed frontier.

It topped Arena.ai's Frontend Code Arena in blind developer testing, it handles text, images, and video natively, and it runs a one-million-token context window. Moonshot built it for long-horizon software work, analyzing large codebases and staying on task across many steps, with a closed-loop system it calls Vision in the Loop where the model reads a screenshot, changes the code, and checks the visible result.

The strategic part is what the weights do to the economics. Through the API, K3 runs $3 per million input tokens and $15 per million output, roughly $0.94 per task, similar to GPT-5.6 Sol and about half of Opus 4.8. Self-hosted, the per-token cost goes to zero and becomes a hardware problem instead. That is the shift everyone is reacting to.

📊 About That 51 Percent

Here is the number driving half the timeline right now. Independent testing by Artificial Analysis put K3's hallucination rate at roughly 51 percent on the AA-Omniscience benchmark, up from 39 percent on K2.6, and that figure appears nowhere in Moonshot's published benchmark charts.

The omission is fair to criticize. The panic is not, and here is why.

First, the metric is not what most people reading the headline assume. Artificial Analysis calculates hallucination rate as incorrect answers divided by all non-correct responses, not as a share of everything the model says. Second, the comparison nobody is posting: Claude Fable 5 scores 54.9 percent on the same benchmark. Anthropic's flagship hallucinates at a higher rate than the model being dunked on today. Third, K3's accuracy went up at the same time, from 33 percent to 46 percent.

Accuracy rising while hallucination rises is the documented signature of standard reinforcement learning training, and it shows up across the industry. K3 got better at answering questions and worse at knowing when it should not. That is a real limitation for fact-sensitive work, it is worth designing around, and it is not evidence that this particular model is uniquely unreliable. Moonshot's technical report publishes alongside the weights today, which should tell us more about how the model was trained.

⚙️ What It Actually Takes to Run This

The weights being free is not the same as the model being cheap, and the practical picture is where the excitement meets reality.

Moonshot's own recommendation is at least 64 accelerators. The full download is reported at roughly 1.4 terabytes even in four-bit quantized form. vLLM needs updates supporting K3's attention architecture before it can serve the model efficiently, and standard prefix caching does not map cleanly onto it. For most teams, that means going through an inference provider rather than self-hosting, at least until the community produces smaller quantizations.

Then there is speed. Artificial Analysis measured K3 generating around 32.6 tokens per second against a 70.9 median for comparable models, and a time to first token of 161 seconds against an industry median of under 3 seconds. For anything interactive, that is disqualifying. For long autonomous runs where you are not watching the cursor, it matters far less. K3 is built for the second thing.

One more caveat worth carrying: Moonshot's coding leaderboard results mix runs from different harnesses, and harness choice alone can move scores by 10 to 26 points on equivalent tasks. Real community replication starts today, now that anyone can download the model and check.

Three days before the weights landed, the most important person in AI hardware posted on X for the first time in his life.

Jensen Huang's debut post was not about GPUs. It was an open letter titled "Open Weights and American AI Leadership," signed by 25 organizations including Nvidia, Microsoft, Meta, OpenAI, IBM, Palantir, Hugging Face, Mistral, Mozilla, and Y Combinator. His framing was direct: AI will transform every industry, power every company, and be built by every country, and open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. Elsewhere he put it more plainly, saying the world needs cutting-edge closed models and cutting-edge open models both.

The letter compares the moment to the 1980s, when software nearly went down a closed path, and asks policymakers to keep the frontier plural by avoiding premature restrictions that would stifle competition or push innovation overseas. Elon Musk, not a signatory, replied that Jensen is right and it has his full support.

The timing was not subtle. Washington is actively weighing curbs on Chinese open-weight models, Kratsios accused Moonshot of distilling a US model to build K3, and the weights were three days out. Worth holding both things at once, though: Nvidia sells the hardware every open model runs on, and it lost roughly $600 billion of market value in a single day in January 2025 when DeepSeek's open release spooked investors. The argument is coherent and Nvidia's commercial interest in it is enormous. Both are true.

The stranger note is who else is claiming this ground. A week before Huang's post, Xi Jinping made his first appearance at Shanghai's World AI Conference and called for a symphony of global cooperation rather than a solo performance by any single country, planting China's flag on openness too. Two superpowers and the world's most valuable chipmaker are now competing to own the same word.

The stories moving fast beyond today's headlines:

Moonshot AI published Kimi K3's full weights at 00:00 UTC today under its own bespoke Kimi K3 License, which is MIT-like but adds commercial conditions above a revenue threshold, making a 2.8-trillion-parameter model the largest open-weight release in history and putting a near-frontier system into anyone's hands. On the Artificial Analysis Intelligence Index it scores 57.1 against Fable 5's 59.9 and GPT-5.6 Sol's 58.9, it topped Arena.ai's blind frontend coding leaderboard, and it runs a one-million-token context with native image and video input. The number dominating the discussion is a 51 percent hallucination rate measured by Artificial Analysis, up from 39 percent on K2.6 and absent from Moonshot's own charts, but the context matters: the metric counts incorrect answers among non-correct responses rather than all output, Claude Fable 5 scores a higher 54.9 percent on the same benchmark, and K3's accuracy rose from 33 to 46 percent at the same time, which is the known signature of standard RL training rather than a defect unique to this model. The practical barriers are real, with Moonshot recommending at least 64 accelerators, a download reported around 1.4 terabytes even quantized, vLLM still needing support for K3's attention architecture, output around 32.6 tokens per second, and a 161-second time to first token that rules out interactive use while mattering far less for long autonomous runs. Moonshot also mixed harnesses in its coding results, where harness choice alone can swing scores 10 to 26 points, so the community replication that starts today is the number that will actually count.

Stay building. 🤖