Meta plays AI catch-up with Llama 4

Artificial intelligence doesn’t pause for catch-up, but that hasn’t stopped Meta from trying. At the company’s inaugural LlamaCon developer conference, hosted virtually on Tuesday, the Facebook owner released Llama 4, the latest version of its large language model that competes with the likes of ChatGPT and Google Gemini.

For the first time, the social media giant has stepped decisively into the generative AI arms race with a clear attempt to close the gap between itself and the companies setting the pace.

Llama 4, Meta’s most advanced AI model to date, includes a standalone Meta AI assistant and a new developer platform, with the latter representing the clearest signal yet that Meta wants to be more than a mere contributor to the world of open-source AI. It wants to be a serious competitor to OpenAI, Google and Anthropic, whose models currently dominate both headlines and enterprise deployments.

For business users, the most important distinction is not that Llama 4 is more powerful, but that Meta is now offering it in ways that make it easier for others to build on top of it, or integrate it into their own products. 

This change in posture may be as significant as the model itself.

A large language model – a system trained on vast amounts of text and code – is capable of tasks familiar by now to most business users: answering questions, summarising documents, writing marketing copy, helping with coding, and assisting in research. With each iteration, these tools become less gimmick and more utility. 

Llama 4 pushes those boundaries further, particularly in its improved reasoning abilities and capacity to operate in multiple languages. It can work with longer pieces of text, understand more complex instructions, and generate more reliable output.

However, benchmarks released since the launch suggest that OpenAI’s GPT-4, Anthropic’s Claude 3 Opus, and Google’s Gemini 1.5 Pro outperform Llama 4 on a range of reasoning and comprehension tasks. Llama 4 is fast, efficient and adaptable, but it is still a second-tier model by technical metrics.

That, however, does not make it irrelevant. Llama 4 demonstrates strengths in specific areas, like its multimodal capabilities, but especially in its cost-effectiveness. 

The truth is that Meta has never been in the race to out-model OpenAI. Its strategy hinges on distribution, openness and community-building, and this is where Llama 4 comes with more teeth. Alongside it, Meta launched a new Llama API that allows developers to access the model via cloud tools, experiment with fine-tuning, and evaluate performance without needing massive hardware investments. For the first time, Meta is offering a full development environment, rather than only the model code.

For developers, this means more freedom to adapt the tool for specific needs. For businesses, it means the possibility of tailoring AI to internal processes without relying entirely on closed systems.

It’s also a strategic move to win trust. Meta has said that data processed through the Llama API will not be used to retrain its models. For industries concerned about data privacy – from law and finance to healthcare – this promise is designed to remove one of the biggest barriers to AI adoption: fear of data leakage.

But for all the talk of openness and access, there is still a catch-up tone to Meta’s efforts. The Llama API arrives more than a year after OpenAI’s developer platform matured, and long after Anthropic and Google launched their full-featured AI portals. Even Meta’s new assistant – being embedded in WhatsApp, Instagram, Messenger and Facebook – may be more of a defensive move than a disruptive one.

That does depend on uptake, but users tended to find the early versions of the Meta AI assistant more of an irritation than a help in WhatsApp. For example, asked on Wednesday to provide information on LlamaCon, it delivered a detailed description of a convention for llama breeders. 

The updated version, which runs on Llama 4 and draws on users’ Meta data to personalise its responses, will be more aggressively pitched as a consumer-facing answer to ChatGPT. It is now being rolled out globally, including to countries like South Africa, where Meta has a significant user base.

By deploying the assistant across its family of apps, Meta is positioning itself as the first tech company to deliver AI at massive scale in daily consumer interactions. Even if Llama 4 isn’t the best model, it might end up being the most used – if it can prove reliable and trusted.

That tension between lagging on model performance and leading on integration was on display in the headline session of LlamaCon, where Meta CEO Mark Zuckerberg shared a virtual stage with Microsoft CEO Satya Nadella. It was a friendly conversation, but one that underlined a power asymmetry.

Microsoft has invested over $10-billion in OpenAI and built its entire enterprise AI strategy around GPT models. Microsoft’s AI crown jewel is Copilot, which runs on OpenAI’s most advanced systems.

While Nadella praised Llama 4’s openness and its availability on Microsoft’s Azure cloud platform, at LlamaCon he came across as supportive rather than committed.

“We want developers to have choice,” he said. “Open-source models like Llama play a critical role in that.”

The irony is that Meta depends on Microsoft’s cloud infrastructure to serve its models to developers at scale. The companies are partners, but not equals. Microsoft leads with its own AI integrations across Office, Teams and Windows. Meta, by contrast, is still in the phase of convincing developers to adopt its tools.

Llama 4, for all its technical competence, is entering a market where switching costs are high and loyalty to existing models is entrenched. Meta’s bet is that openness, privacy guarantees, and community support will offer an appealing alternative.

In some regions, including parts of Africa, this might prove particularly relevant. The ability to deploy AI tools without locking into expensive proprietary platforms could empower local startups, universities and public sector organisations. But this potential depends on training, infrastructure and support – areas where Meta has not historically invested deeply.

Zuckerberg summed it up at the end of LlamaCon: “We’re here to build a more open and collaborative AI ecosystem.” 

* Arthur Goldstuck is CEO of World Wide Worx and editor-in-chief of Gadget.co.za. Follow him on Bluesky on @art2gee.bsky.social.