Meta needs to beat the developers

On Tuesday, she hosts dead The first for her custodian at all Menlo Park, where the company will try to place the developers on the construction of applications with the open Llama Ai models. Just one year ago, this was not a difficult sale.

However, in recent months, Meta has struggled to keep pace with both “open” artificial intelligence laboratories such as Deepseek and closed commercial competitors such as Openai in the advanced AI race. Llamacon comes at a critical moment for the dead in its pursuit of the growing ecological system of Llama.

Winning developers may be simple as charging better open models. But this may be more striking to achieve than it appears.

A promising early start

Dead Llama 4 launch earlier this month Poor developers, with a number of standard degrees in models below like Deepseek’s R1 and V3. It was far from what Llama was: Border model collection.

When Meta launched Llama 3.1 405b last summer, CEO Mark Zuckerberg She described it as a big victory. in Blog postMeta Llama 3.1 405B described as “the most open foundation available”, with the best model for Openai at the time, GPT-4O.

It was an impressive model, certainly – and so the other models were in the Meta’s Llama 3. Jeremy Nixon, who hosted Hackathons in San Francisco Agi House over the past few years, has launched Lama 3 “historical moments.”

It can be said that Llama 3 made Meta loved among the developers of artificial intelligence, as it performed an advanced performance with the freedom to host models wherever they choose. Today, the Meta Llama 3.3 model is often downloaded from Llama 4, said Huging Face’s Product and Growth, Jeff Boudier, in an interview.

In contrast to the reception of the Meta’s Llama 4 family, the difference is blatant. But Lama 4 was controversial from the start.

Correctional

Meta has improved a copy of one of the Llama 4, Llama 4 MAVERICK models, for “Topersality”, which helped her get first in ARNA normative LM. However, Meta has never released this model – the MAVERICK version that was widely released It ended with much worse performance On LM Arena.

The group said behind LM Arena that Meta It should have been “more clear” About contradiction. Ion Stoika, one of the founders of LM Arena, and UC Berkelegor who also participated in the establishment of companies, including Anyscale and Databrics, Techcrunch that the accident had damaged the Developer Community’s Trust in Meta.

“The Meta) should have been more clear that the Mavrick model that was on (LM Arena) was different from the model that was released,” Stoika told Techcrunch in an interview. “When this happens, this is the loss of confidence with society. Of course, they can restore it by launching better models.”

No logic

A blatant neglect of the Lama 4 family was a model of Amnesty International. Thinking models can work carefully with questions before answering them. Last year, many of The artificial intelligence industry released thinking modelsWhich tends to perform better on specific criteria.

Dead Llama 4 thinking modelBut the company did not indicate the date of its expectation.

Nathan Lambert, a AI2 researcher, says that the fact that Meta has not issued a model of thinking with Llama 4 indicates that the company may have rushed to all.

“Everyone calls the thinking model and makes their models look very good,” said Lambert. “Why did you not” define) to wait to do this? I don’t have an answer to this question. It seems as surprising in the regular company. “

Lambert noted that competing open models are more closer to the border than ever and that they now come in more shapes and sizes – which greatly increases the pressure on the definition. For example, on Monday, Alibaba released a set of models, QWEN3Which is alleged to outperform some coding models in Openai and Google on Codeforce, which is a programming standard.

To restore the leadership of the open model, Meta simply needs to provide superior models, according to Ravid Shwartz-ZIV, an artificial intelligence researcher at the Data Science Center at New York University. This may include to bear more risks, such as employing new technologies, he told Techcrunch.

Whether Meta is in a position that is allowed to withstand great risks now is not clear. Tell the current and previous employees previously luck Meta’s Ai Assecing Lab dies slowly. Vice President of Ai Research, Joelle Pineau, This month announced that she was leaving.

Llamacon is a Meta’s opportunity to show what he was cooking to overcome the versions coming from Ai Labs such as Openai, Google, Xai and others. If it fails to deliver, the company may lagward in a very competitive space.