Llama 4, Meta, and China: What’s Next for Open-Source AI
Why we’ll keep getting open-source LLMs no matter how Llama 4 turns out
On April 5th, Meta released Llama 4 calling it “the beginning of a new era”.
It does not seem like it — but that’s ok.
Llama 4 may or may not be bad …
Many users consider the release a flop. On one side there are rumors about Meta “cheating” on benchmarks, on the other it might just be bugs due to a rushed roll-out.
Who knows? There are a lot of different benchmarks and evaluating LLMs is very complex with different performance from task to task. At the end of the day, it’s an open-source model of at least decent quality which is always good for everyone (except their competitors). This leads us to the main point: if it ends up bad, would it impact Meta or the open-source LLM ecosystem?
Why Meta even releases models for free?
When LLM gold rush started with ChatGPT it was Meta who took the open-source route and started releasing their models for free to everyone.
Why would they do that?
To commoditize their complement.
Meta has products that people use, AI complements them => Meta wants AI to be as cheap as possible. Compare that to OpenAI or Anthropic, they sell access to AI, if everyone has it they lose their moat. This is why it makes sense for Meta to release the AI models for free and why OpenAI lobbied for open-source AI to be regulated.
If the goal is to keep the high availability of AI, you need to keep releasing open-source models that closely follow the performance of the state of the art. Meta achieved that with Llama 1, 2, 3. This successfully kick-started the open-source LLM ecosystem with multiple companies releasing their models e.g. Mistral from Europe and DeepSeek from China.
Open source LLMs will keep coming
Currently the DeepSeek models are comparable to the leading closed-source ones in most domains. DeepSeek also keeps releasing new versions without hinting at any plans of switching to closed-source.
OpenAI has after previous hints announced to start open-sourcing some of their models too. (Definitely not the top ones though.)
As a result there are still going to be open-source models with state of the art performance even if Meta stumbles. So sure the Llama 4 release might be disappointing in the short term but that’s not going to affect the long term LLM ecosystem. And who knows — we might get surprised if it’s really just a buggy release after all!
Why is DeepSeek so open?
With China’s DeepSeek pushing the state of the art in open-source LLMs, the question is: will they keep releasing their models for free?
Let’s speculate a bit:
China advancing state of the art in AI has an interesting geopolitical angle. West dominates in software while China leads in manufacturing. If China manages to commoditize the software creation through AI it would benefit them a lot. Therefore it’s in their interest to keep releasing the new models as long as they are not significantly ahead of the West.
Just the DeepSeek models release caused a noticeable S&P 500 drop so the strategy is somewhat working. It’s a bit far fetched though, we’ll see who weathers the storm better. In the meantime DeepSeek seems to be in the good graces of the China’s leadership.
The beginning of a new era
Meta’s Llama models certainly started a new era in freely available AI successfully pushing the price close to zero. Similarly DeepSeek has started a new era with the rise of China as a main open source contributor.
As a result the whole world has access to highly capable AI for free. The future has never been so evenly distributed 👏