--- title: "The Door Is Open, but the Recipe Is Still Locked in the Drawer" date: 2026-11-05 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-open-weights-05-1105b language: en --- # The Door Is Open, but the Recipe Is Still Locked in the Drawer > From November 4 to 5, Moonshot AI put a large model called Kimi K3.1 Pro online in full. In the past, using it meant sending questions through the interface the company had opened and waiting for an From November 4 to 5, Moonshot AI put a large model called Kimi K3.1 Pro online in full. In the past, using it meant sending questions through the interface the company had opened and waiting for an answer. This time was different: the key numbers that let the model read text, recognize images, and write code could be downloaded by anyone and taken home to tinker with at leisure. For ordinary people who do not work with technology, the whole thing can be put in kitchen language. To eat this dish before, you ordered from a menu at a restaurant, and the kitchen was closed to outsiders. Now the shop sets out a whole finished pot and lets you carry it home to reheat and season yourself. What you carry away is the finished dish, not the craft behind the stove. Do not underestimate that pot. According to the official introduction, it uses a mixture-of-experts architecture, with 2.8 trillion total parameters, about 104 billion actually active in a single inference, and the ability to take in material as long as 1 million tokens at once, as well as a native knack for understanding images. Keep these numbers roughly in mind, because they will come up again. Nor is it a brand-new foundation trained from scratch. According to the official account relayed by The Paper, it is an upgrade refined bit by bit on top of the original K3 base through large-scale joint reinforcement learning across many domains. The foundation did not change; the interior was renovated again and again, and the load-bearing frame is still the original one. ## A Box of Parameters Comes Home With the weights in hand, the most concrete change is that initiative changes hands. Companies and developers no longer have to queue for a reply from an interface. As long as they can afford the computing power, they can install the model in their own machine rooms, keep the data from leaving, and then adjust and modify it to suit their industry, so it fits their own pace of business better. That is a different matter from borrowing something you may not alter. In the old interface days, how the model reasoned and how its parameters were arranged were things users could never touch, like ordering across a counter. What arrives now is the skeleton fixed at the end of training; you can fine-tune it further, or take it apart to see exactly what it has remembered. According to The Paper, the news drew plenty of discussion in overseas technical communities, and the official NVIDIA AI account also posted public congratulations. That kind of congratulation speaks to how lively the moment was, but it does not change one thing: what can be carried home has always been only this skeleton, not the method for making the skeleton. ## The Words Open Source Carry More Weight This is exactly where misreadings come easiest. Many retellings take the shortcut of calling it open source, but the precise term is open weights. The Paper points this out specifically in its report: opening model weights this time is not the same as traditional open source as defined by OSI, and the difference of two words leaves a wide gap between them. OSI is the Open Source Initiative, and its old rule for open source is quite plain: it is not enough to let you use something; the source code and the whole process of reproducing it from scratch must be laid open too, so that anyone following along can build the same thing. Applied to a model, merely handing over the finished product is still a long way from that line. One way to remember it: open source is like being handed the recipe, the list of ingredients, and every step of the heat, so you can cook the same pot again on your own stove. Open weights is like being handed a pot already cooked, free to reheat and season as you like, with no word on how the flavor was adjusted spoon by spoon in the first place. ## The Recipe and the Ingredients Were Not Included In the wording of The Paper, this time neither the training code nor the original training datasets were made public. The training code governs how the model is taught step by step; the original data governs what it is taught with. Together the two make up the complete recipe for reproducing the dish from scratch, and missing either one leaves it incomplete. Without those two, outsiders can see clearly what the finished product looks like, but cannot work backward to why it turned out that way, still less cook a new pot with a different batch of ingredients. Being able to download, deploy, and fine-tune, and being able to build one from scratch, are three entirely different things, with that whole method standing between them. That also explains why people in the field choose their words so carefully. Weights are the result; training code and data are the process. Handing over the result but not the process has the advantage that others can use it right away, and the cost that no one can retrace the path for you. A difference of one word in the name makes a wide difference in what you are allowed to do. ## Commercial Use Still Depends on the License There is one more door not yet mentioned: what you may do after obtaining it depends on the license. According to 21st Century Business Herald, what was published alongside the weights is a license agreement drawn up by Moonshot AI itself, tagged as a non-standard type on the code-hosting platform rather than one of the familiar general open-source licenses. The agreement is not stingy. Use, modification, deployment, fine-tuning, redistribution, and even sale are all permitted. Its restrictions sit mainly on commercial use at scale: for example, packaging the model into a service resold widely to outsiders, or building a product with a very large user base, calls for a separate agreement or a visible credit noting the source. For ordinary developers and companies that keep the model for internal use, most of the terms are generous enough that there is rarely any need to negotiate day to day. The parties who really sit down for detailed talks are those planning to do large-scale business with it to the outside world. The door is open, but the threshold is written on a sheet of paper that has to be read clause by clause. ## The Hardware Barrier Did Not Disappear Nor should anyone assume that once weights are released, everyone can run the model on a home computer. In the wording of The Paper, a very large mixture-of-experts model at the 2.8 trillion level carries a high hardware deployment barrier of its own; simply fitting all those parameters steadily into memory is more than an ordinary machine can handle. The more realistic route is still the interface and cloud services. What the weights open up is the upper limit of choice: institutions with the means can keep the model inside their own network so data need not flow out; those without still call it by usage. For most people, what changed over these two days is not the device in their hands but one more option they can discuss. Separate the two questions and it becomes clear. Whether you can obtain it is a matter of license and download; whether you can carry its weight is a matter of machine rooms and budgets. The first did loosen over these two days; the second did not get any lighter along with it. Blur them into a single claim that everyone can use it, and people without a rack may be set up for a hollow joy. ## Getting the Name Right Matters Most So the thing most worth remembering in this news is not that string of daunting parameters, but the difference between two terms. Open weights says what you are allowed to take; open source says how much the other side has handed over. One is about the finished product, the other about laying the method bare alongside it, and they were never measured on the same scale. The next time you see a model described as open source, try asking a few questions first: is it the weights that are open, or are the training code and original data given too? Can you deploy it yourself, or only use it under the rules as written? Can you change it, or only reheat it as it came? Once those questions are answered, the excitement settles back onto facts that can be checked. For ordinary people the lesson is not far off. A dish arriving at the table is certainly good news, and being able to carry it home to eat slowly is better still. But if someone casually tells you the recipe came with it too, you would do well to reach over and pull the drawer open for a look, because that compartment, this time, is still locked. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-open-weights-05-1105b Free to quote with attribution and a link to the original.