Transcript 0:00 This week has shaken our understanding of how far an open source model can go. Two models launched in the same week, one American and one Chinese, which often means battle. 0:11 Both are receiving praise from the open source community, yet they could hardly be more different. One is deliberately not presented as the best. The other is beating Fable-05 and GPT-5.6 Sol in the front-end code arena. 0:26 Analyzed together, what do these two models tell us about where the industry is going? [upbeat music] Attention Span. Welcome to Attention Span by Turing Post. 0:39 I'm Ksenia, and today, we're going to examine these two strangers, why both models are so enormous, and what kind of control open weights provide when very few people can run them. 0:52 The American model is Inkling, the first major model from Mira Murati's Thinking Machines lab. Mira Murati used to be a CTO of OpenAI. The model has 975 billion parameters. 1:05 It's a big model, and its weights are available to download right now. Thinking Machines says openly that the stronger models exist. 1:13 The model is intentionally not the best out there, but the company is arguing that the ability to customize a model can count for as much as its initial benchmark position. The Chinese model is Kimi K3 from Moonshot AI. 1:30 It has 2.8 trillion parameters, a huge model. The previous Kimi model ranked 18th in the front-end code arena. K3 entered at number one, ahead of Claude Fable 5. So this is not a simple comparison between two models. 1:46 Kimi is testing how close an open model can get to the capability frontier. Thinking Machine is testing how much value can be created by adapting a model after it has been trained. A couple of words about being open. 2:01 I used open source at the beginning because that is how these models are often discussed. For AI in general, open weight is the more accurate term. An open weight release gives you the model's learned parameters. 2:14 You can host the model, compress it, fine-tune it, and build a service around it. 2:19 It does not necessarily include the training data, the full training code, or everything required to reproduce the model from the beginning. 2:26 But using open weights allows you to fine-tune it and use it for your own needs. So Inkling is available now, and Thinking Machines released the weights under Apache 2.0, together with an acceptable use policy. 2:41 Kimi K3's API is available. As I record this, its weights are not yet published. Moonshot says the full weights will be released by July 27th, together with more technical details. 2:54 So we have not yet seen the final license. Let's begin with Kimi because its performance attracted the most attention. It blew up on Twitter, as traditionally happens on Twitter, now known as X. 3:07 And when that happens, the servers often go down. So if you're unable to try Kimi K3 in the browser, wait a little. Something else will be released soon, and everyone's attention will switch to that. 3:19 Anyway, what happened and why is it so hot? Kimi K3 shot up from 18th place to number one in the front-end code arena. 3:29 It scored 1,679 points and ranked first in six of seven categories, including consumer products, data and analytics, simulations, marketing, and reference-based design. It came second in gaming behind Fable 5. 3:44 The front-end code arena uses blinded human comparisons. People look at two generated interfaces and choose the one they prefer without knowing which model produced it. 3:55 This is where the claim that K3 is beating Fable 5 and GPT-5.6 Sol comes from. It is a strong result, and it also a specific result. It tells us that people preferred K3's front-end work. 4:10 It does not establish K3 as the best coding model for every task. On the broader Artificial Analysis Intelligence Index, K3 scores 57. Fable 5 scores 60, while the strongest GPT-5.6 Sol configurations score 58 or 59. 4:28 Moonshot itself says that K3 still trails the strongest proprietary models overall. 4:34 So K3 is not winning everywhere, and it's still absolutely extraordinary because a model scheduled for an open weight release can now compete directly with the strongest closed models on visible, useful work. 4:49 Now, to the enormous number and how useful can this model can or cannot be for you. The model has 2.8 trillion parameters. K3 is a mixture of experts model. It contains 896 expert modules. 5:04 These are groups of parameters that develop different specializations during training. For each token, a router selects 16 of those 896 experts. 16 out of 896. Why? 5:19 Well, this reduces computation because K3 does not use all 2.8 trillion parameters for every token. Oh, that's why. 5:28 It does not reduce the storage requirement in the same way, though, and all 896 experts still need to be available because the router can select a different group for the next token. What does this mean for you? 5:41 If you have a regular Mac or one personal GPU, you cannot run Kimi K3 locally. Even compressed to four bits, the weights alone need roughly 1.4 terabytes of memory. 5:53 A powerful consumer GPU has tens of gigabytes, not thousands. Moonshot recommends at least 64 accelerators for efficient deployment, so K3 becoming open weight will not turn into a local model. 6:08 Most people will still use it, it through a website, an API, or a cloud provider. The practical benefit is that Moonshot may no longer be the only company allowed to host it. 6:20 Other providers can run the same model, optimize it, compete on price, and offer different privacy or deployment options. For a regular user, that competition is the useful part of the open release. 6:33 K3 is also not especially cheap by open model standards. Moonshot charges $3 per million on cached input tokens, and 15 per million output tokens. Cache at input costs 30 cents per million tokens. 6:48 Another interesting moment is that price per token does not tell you the full cost. Artificial analysis reports that K3 generated 130 million output tokens across its intelligence index evaluation. 7:02 The median model in its comparison group generated 63 million. Longer reasoning can make a model more expensive even when the listed token price looks competitive. 7:12 This is a change from the familiar open model proposition. K3 is not primarily offering a smaller and cheaper alternative. What it's offering is high capability with the possibility of independent host. 7:26 Once the weights and license are available, Moonshot will no longer be the only organization able to operate K3. 7:34 Another cloud provider can host it, a government can deploy it under its own operational rules, a company can choose a provider based on location, privacy requirements, performance, or price. 7:46 Will American government use it? Hmm, doubtful, but wait for the next part. It actually might be the case also. So most users will never operate 64 accelerators that Moonshot suggests for this model. 7:59 They can still benefit if several companies are allowed to serve the same model. With a closed model, the developer usually controls the model, the API, the price, and the hosting environment. 8:11 Open weights allow these functions to be separated. One company can train the model, another can host it, and a third can adapt it for a customer. This is the control that Kimi may provide. 8:23 It is control for operators first, with indirect benefits for everyone who uses their service. Now to Inkling, because there is interesting part about Inkling and other service that Thinking Machines created. 8:36 So let's look at Inkling. Inkling has 975 billion total parameters, with 41 billion active for each token. 8:45 It accepts text, images, and audio, and a lot of people praise it for audio, and supports a context window of up to 1 million tokens. 8:56 On the artificial analysis intelligence index, Inkling scores 41, compared with K3 at 57. Thinking Machines says directly, "Inkling is not the strongest overall model available today, open or closed." 9:10 That is an unusual way to introduce a model after spending a great deal of money and time to train it. But Inkling makes more sense when you look at Tinker. Tinker is Thinking Machine's training platform. 9:24 It allows customers to fine-tune models while Thinking Machines manages the distributed infrastructure. Tinker uses LoRA, or low-rank adaptation. 9:34 Instead of updating every parameter in the base model, LoRA trains a much smaller set of additional weights. The base model stays largely unchanged while the adapter learns a particular task, style, workflow, or domain. 9:49 This is much cheaper than retraining a giant model. It still requires good examples and reliable evaluation. If an organization cannot define the task or measure the result, fine-tuning will not solve the problem. 10:02 When task is clear and the organization has useful data, a customized model can outperform a stronger general model on that specific work. 10:12 In my conversations with other companies, I know that the companies that need precise work with their specific data, they use local models specifically for that. 10:23 And here's the important business detail from Thinking Machines. Tinker does not support only Inkling, so this is where Kimi K3 might enter at some point. 10:34 Customers who use Tinker can also fine-tune models from Kimi, DeepSeek, and IDEA, OpenAI, and Qwen, and whatnot. 10:42 Thinking Machines is therefore competing in the model market while selling a platform that can benefit when another model performs better. If Inkling is the right base model, Tinker can customize Inkling. 10:55 If Kimi is better for the task, Tinker can customize Kimi. 10:59 Inkling does not need to lead every benchmark for the business to work, and talking about the US government, if they trust Tinker and its safety mechanism, it can use Kimi. Through the platform if they want. 11:12 This explains why Thinking Machines is comfortable saying that Inkling is not the best model. The company is building around the ability to change models, including models it did not train. 11:24 Inkling is much smaller than K-3, but running it still requires substantial hardware. The full precision checkpoint needs at least two terabytes of combined GPU memory. The official MVFP4 version needs 600 gigabytes. 11:41 Within a day of release, Unsloth published community 1B quantizations at roughly 270 and 285 gigabytes. 11:50 These versions can run on a machine with about 290 gigabytes of system memory, although Unsloth reports roughly 74% accuracy retention. 12:00 The quality reduction is substantial, but the speed of the community's response is still important. 12:06 Thinking Machines released the official checkpoints, and other developers immediately explored how far the hardware requirements could be reduced. This is one practical advantage of open weights. 12:17 The original developer does not have to perform every optimization or support every type of hardware. Other groups can experiment with compression, serving software, and fine-tuning methods. 12:29 So why are we talking about Inkling and K-3 together? 12:32 Not only because they were launched the same week, but also to explore and follow the scientific cross-pollination between companies and countries, because we heard a lot from Anthropic specifically how Chinese companies steal and distill their models. 12:52 But let's look how Inkling was created, and that it's absolutely normal to work with other companies, research, and innovation. So Inkling and K-3, as we discussed before, represent different strategies. 13:08 Their development is already connected. Inkling's mixture-of-experts architecture largely follows DeepSeek-V3. 13:15 Thinking Machines also says that a small part of Inkling's early supervised fine-tuning used synthetic data from open weight models, including Kimi K 2.5. 13:26 An earlier Kimi therefore helped train a model that now competes in the same market. There is an irony there that an American company uses data from a Chinese model, don't you think? Or maybe it's just absolutely normal. 13:39 And if we talk about Kimi K-3, there is another interesting part about it. Moonshot says that an early version of K-3 handled most of the team's GPU kernel optimization work late in development. 13:51 The engineers defined the tasks, supplied the environment, and reviewed the results. K-3 helped optimize some of the software used to run the final model. 14:01 There was a bunch of jokes on Twitter about Kimi K-3 distilling the American model that does not yet exist. Even those examples do not prove autonomous recursive self-improvement. 14:13 These are controlled engineering processes. They do show that models are becoming useful tools inside model development, both in Inkling and on Kimi K-3. 14:23 Open weights make direct reuse easier, and that's the beauty of it. Other labs can fine-tune the checkpoint, generate synthetic data, study the architecture, or develop new inference methods around it. 14:36 A model can therefore serve users and contribute to the development of later models, including models from other companies and other countries. None of this makes closed models obsolete, of course. 14:49 Fable-5 and the strongest GPT-5.6 Sol configurations remain ahead of K-3 on Porter evaluations. 14:56 Closed providers also manage deployment, updates, security controls, support and integrations, and communication with the government. 15:05 For many companies, paying for an API is more practical than operating 64 accelerators or maintaining a specialized inference team. Commercial API can also include contractual privacy and retention terms. 15:18 And using a closed model does not automatically mean that customer data is used for training. What Kimi K-3 does is that it gives customers a stronger basis for comparing the premium. 15:31 Are they paying for higher performance of the, their work, better reliability, faster responses, stronger support, or lower operating complexity? If the provider delivers those benefits, the premium can be justified. 15:44 What do these models tell us together? Kimi K-3 and Inkling show that open models are developing in more than one direction. I think it's a very important moment. 15:54 Kimi is pushing open weights toward higher general capability, even when the model requires data center scale infrastructure. One of the employees of Kimi's, of Moonshot said they are religiously achieving AGI. 16:07 Thinking Machines, on the other hand, is building around customization. Everything about them is very practical. Inkling gives the company its own base model, while Tinker can adopt models from several competing labs. 16:20 Open models are therefore no longer defined mainly as smaller and cheaper alternative to closed systems. This paradigm is gone. 16:28 An open model can be enormous, expensive to operate, and competitive with leading closed models. Another can begin from a lower benchmark position and create value through adaptation. The shared benefit is choice. 16:42 More organizations can decide where a model runs, who operates it, how it is adopted, and which provider serves it. 16:50 Kimi K-3 and Inkling both show how cross-pollination, how collaboration and exploring other companies' and countries' innovation can improve both sides. The limitation is infrastructure. 17:05 Access to the weights does not provide the hardware, software, data, evaluations, or expertise required to operate them well. K-3 suggests that an open model can approach the capability frontier. 17:16 Inkling suggests that the initial model score may become less important once organizations can customize the system around their own work. 17:24 Together, they show an industry competing across several layers: model quality, operating costs, hosting, customization, and the software used to manage all of it. This is where open source and open weight AI is going. 17:38 The model remains important alongside the infrastructure and customization systems built around it. That's it for today. 17:46 Tell me in the comments which strategy you find more convincing, Kimi's focus on capability or Thinking Machine's focus on customization. And what do you think about this new development for open weight, open source AI? 17:57 Thank you for watching, and I'll see you next week. [upbeat electronic music]