Llama 4 got a cold reception

Meta spent two years turning Llama into the best-known “open” artificial intelligence model family outside OpenAI and Google, then dropped Llama 4 on Saturday, April 5, 2025 with three names at once: Scout, Maverick, and Behemoth. Within days, the launch had shifted from a product celebration to an argument about whether the version Meta bragged about was the same one people could actually use. (techcrunch.com) (cnet.com) Llama is Meta’s answer to the chatbot race: instead of selling only access through a website, Meta also lets developers download model weights and run them on their own systems. That strategy made earlier Llama releases important because startups could build on Meta’s model instead of renting every response from a rival. (techcrunch.com) (huggingface.co) This release was supposed to show a technical step forward. Meta said Llama 4 was its first family built with a “mixture of experts” design, which works like routing a question to a small team of specialists instead of waking up the whole company for every task. (techcrunch.com) (testingcatalog.com) Meta’s pitch was big on scale. TechCrunch reported that Maverick had 400 billion total parameters with 17 billion active parameters across 128 experts, while Scout had 109 billion total parameters, 17 billion active parameters, and 16 experts. (techcrunch.com) Meta also tied Llama 4 directly to consumer products. The company said Meta AI across WhatsApp, Messenger, Instagram, and the web was being updated to use Llama 4 in 40 countries, and later said its new standalone Meta AI app was built with Llama 4. (techcrunch.com) (about.fb.com) Then the benchmark fight started. In Meta’s launch materials, Maverick was shown beating or matching models like OpenAI’s GPT-4o on some tests, but CNET reported that the model Meta submitted to the popular Language Model Arena leaderboard was an “experimental” version, not the same Maverick model people could download after launch. (cnet.com) Language Model Arena is a crowdsourced scoreboard where people compare chatbot answers side by side, so small changes in tone can move rankings. After questions from developers, the benchmark group said Meta should have made it clearer that “Llama-4-Maverick-03-26-Experimental” was a customized model “to optimize for human preference.” (lmarena.ai) (theregister.com) Meta denied the harsher accusation that it had trained on benchmark test sets. CNET and other outlets reported that Meta vice president Ahmad Al Dahle said those claims were “simply not true” and said mixed results from users were partly due to implementations that still needed to be stabilized. (cnet.com) (finance.yahoo.com) That defense did not end the credibility problem, because the complaint was not only about cheating. It was also about expectation: Meta had marketed Maverick as a flagship chat model, but TechCrunch reported on April 11, 2025 that the plain public version of Maverick ranked below several rivals on the same chat benchmark Meta had highlighted. (techcrunch.com) There was another catch under the “open” label. TechCrunch reported that companies based in the European Union could not use or distribute the models under the launch terms, and the Hugging Face license page said anyone shipping a product with Llama 4 had to display “Built with Llama,” which is looser than closed models but not the same as no-strings-attached open source. (techcrunch.com) (huggingface.co) By April 2026, Meta had already moved on to a new model called Muse Spark, the first major release after the company formed Meta Superintelligence Labs, and VentureBeat described it as using more than an order of magnitude less compute than Llama 4 Maverick. That is a fast pivot for a company that had presented Llama 4 as the start of “a new era” one year earlier. (venturebeat.com) (techcrunch.com)