Meta is in a tough spot with Llama
Notes on Llama and Llamacon
Hello, and welcome to the seventh edition of the AI Report. We aim to inform you about trends in the world of AI—from research to products to everyday life—and shed light on the latest trends. Please subscribe for actionable insights and share this newsletter on social media.
Trends in AI: Focus on Llama
Llamacon 2025 was on April 29th, and it was underwhelming.
No reasoning models
The biggest disappointment of Llamacon is that there was no announcement of a reasoning model. OpenAI and Google already had reasoning models last year. DeepSeek-R1 was announced in Jan 2025. Qwen has released Qwen3 (one day before Llamacon!) which can reason (in fact, Qwen's previous QwQ-32B is not very good by modern standards, but it was released in 2024). Microsoft announced Phi-4 models a few days back. In fact, a small group of talented individuals announced OpenThinker models in January of this year.
Llama 4 was announced a few weeks back, without any reasoning model, and Llamacon came and went.
And no word from Meta yet.
Llama API: Cannibalize the partners
Meta announced that it will support inference and finetuning, Together.ai's two bread-and-butter products. Companies like Cerebras are also deep partners of Meta since they serve a lot of llama tokens.
However, it is unclear how serious Meta is with this new product.
As anyone who has read about strategy will say, you must aspire to win. So if Meta doesn't aspire to win, it will be a mid-product (can you depend on it for your production use cases?). But if they want to win, they are going to cannibalize the partners.
So, what are Together and Cerebras to do now? They will want more adoption of DeepSeek and Qwen, but those are not American models.
What was Satya Nadella doing there?
Satya Nadella flew to the event, which is kind of interesting to me. Since Satya wants to reduce reliance on OpenAI and Llama is trying to increase its adoption, we will probably see a deal go through between them. We could see Llama models show up more prominently on Azure to displace some of OpenAI. But can llama models displace OpenAI? So, I suppose Microsoft might invest some money into Llama training?
Llama 4's mixed landing
Llama4 models had a mixed landing. First, nobody knows why Meta announced the models on a Saturday afternoon when everyone usually takes a nap.
Perhaps they want to show they are cool. Or possibly Zuck had a firm deadline and someone didn't want to get fired.
The other fiasco was that they put out one model on HF and a slightly different one (which is kind of sycophantic) on Lmsys leaderboard to rank higher on the leaderboard.
Beyond having hard-to-remember names, these Llama 4 models operate in a funny zone. They are all mixtures of experts, which means they are really built just for inference. But better models exist!
Llama3 models were shining, especially the 8B version, because they were easy to finetune and customizable. The new mixture of expert models is more complex to finetune, and thus, it has lost the appeal of the developer crowd who want to customize models.
Oh, and the Llama 3 8B was perfect for running on the MacBook. The Llama 4 models don't fit very well.
So what gives?
Meta is in a tough spot. The Chinese are outcompeting Meta, and at the same time, Meta seems to have made a series of mistakes with llama 4 launch and now announcing the API. Their only hope is now if they create a very strong reasoning model and open-source it. Meanwhile, we wait.