Wide Open
China's Kimi K3 open weight model has shaken Silicon Valley and even given rise to conspiracy theories but are open-source LLMs really a threat to national security?
Over the past couple weeks China has been making waves from across the Pacific by launching a series of open-weight LLMs that appear, at least at first, to be just as powerful as the frontier models coming from Anthropic and OpenAI. This deluge of cheap and powerful tech from a rival nation has freaked out everyone from the leaders of the aforementioned companies, to their investors, to leaders in Washington, the conversation at times culminating in calls to “ban open-weight AI”, a practice that would be deeply anti-competitive and short-sighted when it comes to innovation in the AI space. These models also launch on the heels of accusations from Anthropic about some Chinese hackers attempting to gain intel on their models.
It’s no secret that America and China have been involved in something of a cold war. If it’s not a Cold War, then it’s a lukewarm war getting colder as China continues to grow its power and influence and as America continues to isolate itself on the world stage. The question everyone should be asking though is, are these actions inherently adversarial or is it just a sign that AI is gaining widespread popularity and acceptance?
As we dig into this topic I want to clarify the definition of open-weight models such as DeepSeek, Mistral, Kimi, Qwen, and Llama. While these models are often called open-source and “spiritually” capture the intent of open-source by making themselves freely available to fine-tune, they typically do not release the full training datasets or pipelines that would be required to satisfy a traditional open-source definition.
Open and Shut Cases
The core issue around these open-weight models is that they are, well, free. In contrast there has been some serious concern around the business models of both Anthropic and OpenAI charging exorbitant prices for seemingly arbitrary token processing; the bills for which have been eye-watering for business leaders. Open-weight models on the other hand simply need to be installed on a capable machine or server where the tokens can be processed locally, without the processing fees charged by the frontier models.
The benefits of running an open-weight model seem obvious at first, a user gets access to a powerful model for pennies on the dollar compared to signing a contract with Anthropic or OpenAI. There’s a catch though, it’s not plug-and-play, necessarily. Installing a model and running it requires the correct GPU processing and environment, so having someone with the technical know-how to manage this infrastructure is a must. That said, the savings are significant.
All that said, it needs to be stated how much scaffolding and ML support both Anthropic and OpenAI have built around their products. Up until now, these companies have hidden these apps/functions behind the curtain of their LLM interfaces to provide a bit of a Wizard of Oz effect. When a user needs to perform a task the LLM calls these functions and presto, it magically happens.
That’s not the LLM doing that magical thing, that’s a function on the backend, and that may be the thing that saves OpenAI and Anthropic from this new wave of competition at the end of the day.
No so Super Models
What these Chinese companies have effectively commoditized tokens. They’ve said “We’ve made tokens worthless now you have to compete with us.” That’s fine, the token business model was terrible and it was coming apart at the seams. There is no modern computing business that works like that, where you’re nickel-and-dimed for usage. The model reminds me of the long-distance calling business model where the caller paid by the minute. The internet decimated that model since that traffic moved to email, video calls and other modes of communication.
What Chinese models do is embrace edge computing, which is definitely an existential threat to Anthropic, OpenAI and the data center build out if the entirety of the processing happens on the device. The main issue with this approach is that installing models locally is difficult for the average person to do, and models are still too large to be installed on mobile devices, even the smaller variety like 1B or 3B models.
Eventually we may see these models appear on android devices by default, Apple would likely have their own frontier model that would be loaded on-device. The size needs to be considered since a 16B model can consume 32 GB of storage on a device. Even the smaller models can consume up to 2.5-6GB of storage. The size of the model matters because the more storage dedicated to the model, the less storage for other Apps, photos and videos. Whether these models would get regular updates and how that would work is another question. The other consideration is the processing power. The newer iPhones have no problem with processing, the beefy A19 chip could handle the duties of a local model. Lower-end Android devices might struggle though. In both cases though, this approach would improve performance dramatically but the models would still need to make calls out beyond the device to achieve more impressive feats.
Embracing edge computing makes the models have near-zero latency, that’s a big problem for the frontier models built as services because they have to transmit data to a data center and back to you the user. Again though, the hardware and the use case aren’t there yet, but it will be very soon so this is something Anthropic and OpenAI need to solve for now.
To be perfectly frank the business models for the American frontier models were questionable to begin with, charging for tokens and processing speed isn’t a value proposition, it’s a technical specification and it’s easily replicable, as evidenced by this recent wave. What’s still necessary, even for models running on the edge is the need to connect to the backend somewhere, and this is where OpenAI and Anthropic can excel. These companies have rounded out their products and offerings, and that’s where the real value lies; not in the models themselves.
A Threat to the Duopoly
The frontier models have gained a reputation for being inhospitable to start-ups and others trying to develop the AI space alongside them. Sam Altman infamously claimed that OpenAI will simply “steamroll” anyone trying to build a product around ChatGPT, the great irony of this being that Altman came from the startup incubator Y Combinator, and that statement is the polar opposite of incubation.
As a founder of a nascent start-up myself, I quickly dismissed both ChatGPT and Claude as viable options for integrating into a product. There were several reasons for this and those reasons have only become more pronounced as time has worn on. Altman’s quote was a reason for pause, but it only reaffirmed that OpenAI is building a product for themselves, not a platform.
As new versions were released of both Claude and ChatGPT, the products would add and remove features. It’s impossible to build on that, it’s like building a house on shifting sands.
Additionally, there are privacy concerns, where the data is going, who owns it and how it’s being handled. There’s also the lack of ownership over the tech itself as well, and the inability to have some say in what it does or doesn’t do, and the outputs, where the data is stored, who owns what, it’s a mess.
But the real deal breaker was cost. Not only is being “GPT/Claude wrapper” wholly uninteresting but it’s an untenable business model, it doesn’t scale. The overhead for paying by the token is insane. If your users process too many tokens you’re out of business instantly. The vast majority of the revenue would be lost to the foundational models, simply for performing basic tasks.
As a result I opted for open-weight to preserve autonomy and optionality. Since starting Group Table the leader of the AI space has changed a half dozen times and now we have this new influx of Open-Source models entering the fray. That said, open-weight models aren’t a new concept in this space, far from it.
Opening the Floodgates
As ChatGPT, Claude and Gemini duked it out for “dominance” in the emerging AI market, open-weight models fell by the wayside, but they never fully went away. I’ve been critical of Meta, mainly for their approach to leadership and company culture but they got one thing right, when they made Llama Open-weight, they were way ahead of the curve. If only they had the foresight to stick with it, they might have been able to capitalize while their competitors bloodied each other up.
The last decent Llama model was Llama3 which is still fine in terms of its benchmarks, the benchmarks are a bit performative and a lot of it is really just shuffling stuff around. When companies release their results of how their models perform against benchmarks, these are tests that LLMs are designed to excel at and the models are optimized to do well on the tests, so it’s not like these are wholly real world scenarios.
There have been major innovations to LLMs since Llama3 was introduced but those have been more with regards to memory and storage which are more about scaffolding and data management than the model itself. Another open weight model worth mentioning is the French frontier model Mistral, which doesn’t get enough notoriety, but it keeps advancing unlike lLama.
All this to say is that this deluge of Chinese models is new, but it’s not novel. Those working with LLMs have been working with open-weight models for years but the notion that China has somehow coordinated an “attack” has grabbed headlines and led to some dubious assertions on how the U.S. should respond, if at all.
Panic in San Francisco
To say people are panicking about these open-weight models from China would be an understatement. It’s par for the course with AI, it seems that the temperature is cranked up to high on literally everything and any news causes a massive freak out. The launch of these models has even stoked conspiracy theories. Professor Scott Galloway, whom I only know as an internet pundit, was asked about this for some reason and decided to offer his opinion on how it would crash the stock market/economy.
Professor Galloway opines on business but he’s a bit out of his depth here. His assertion is that China is flooding the market with “Good enough” models to undercut the AI sector here in America where there has been massive investment.
Ok, where’s the problem? Neither OpenAI or Anthropic are public and Nvidia is still selling chips. Cheaper models still need compute, which is where all this investment is going. It's going into buying chips and building data centers. If these open-source models lead to more AI adoption that should increase the need for data processing, not decrease it.
The only existential threat is if the models need less compute in which case we have a surplus of compute no one wants, it crashes the price of compute and the Googles, Metas, Oracles of the world lose their shirts on the bet they made building out capacity we didn’t need. That’s the free market for you.
The problem with conspiracy theories like Professor Galloway’s is that it leads to panic and with panic comes crack downs. Already we’re hearing from Washington murmurs of open-weight bans, which would kneecap America’s ability to compete. Jensen Huang and a number of tech leaders have come out signing a letter in favor of open-weight models because it’s not the apocalyptic scenario Galloway suggested, it would increase competition and usage of chips, something this sector needs to remain afloat.
Has China Closed the Gap?
Many are pointing to this wave in hyperbolic terms, claiming that America’s lead has eroded and that China has closed the gap or taken the lead. From a technical perspective China has just introduced more models into the marketplace to try to take market share. This has long been a hallmark of Chinese industrial strategy: competing aggressively on price and scale.
From a technical perspective there was no gap to close. LLMs have been on a virtual plateau for a few years, the major improvements being to data retrieval and memory. Again, this pertains more to the scaffolding being built around the products than the actual LLMs themselves. So a user could feasibly build a better memory retrieval system than OpenAI or Anthropic but the question would be if it would be scalable in the same ways?
That brings us back to the earlier point. Right now Claude and ChatGPT are more than just LLMs. The ecosystem they've built around those models is what makes them what they are, not necessarily the models themselves. So in this sense, no, China hasn't caught up. Unless these open-weight companies are willing to throw a ton of resources at rebuilding everything that surrounds Claude and ChatGPT, they won't be the same. That's the moat for these companies at present, and it might behoove them to double down on what has become their specialty.
For the Open-weight models, it doesn’t really make sense to add that overhead of maintaining a bunch of applications and functions. That’s why the models are open-weight, so the users can do that themselves. The high bar of technical ability makes it clear who the audience for these models are, IT professionals, Software Engineers or some blend of the two. There have been jokes about companies wanting to build their own LLM, well with open-weight, they essentially can. The question is if free now means free forever?
Trojan Horses
Everybody loves free stuff. A COO and CFO reviewing the bill from Anthropic would jump at the chance to replace that expense with something at a fraction of the cost. We all know nothing in life is truly free though.
One major question for western companies with any foreign technology, especially from a rival like China, is if there are hidden security risks. The other is more practical, the model itself in the case of Kimi might just be a loss leader, similar to Mistral. Give away the model and then sell the enterprise capabilities later.
At the moment though this wave seems like a trial balloon more than a full-on invasion by the Chinese frontier models. If there’s a larger end game at play, only time will tell.
The Future is Open
Despite the objections of Dario Amodei and some xenophobic pundits. Open-weight models are good for innovation, it’s good for the AI space and good for the substantial investment being made in AI. It is a threat to the stable of frontier models that have established themselves here in the US but only because that was an artificial environment created by a near monopolistic country club of companies.
If we’re to see AI truly grow and explode into what the internet became and what Apps became after the introduction of the iPhone, then we need openness, we need for every engineer to be able to pick up a model and tinker with it. We need for start-ups to be able to grow without getting crushed in the machinery of big tech.
While I’m not necessarily an advocate for these models coming from China per se, I am an ardent advocate for open-weight models and America needs to step its game up. There should be more models and innovation coming from the U.S. not just a handful of companies and the resurgence of open-weight can be an exciting new start.