Meta has introduced Llama 3, the next generation of its state-of-the-art open source large language model (LLM). The tech giant claims Llama 3 establishes new performance benchmarks, surpassing previous industry-leading models like GPT-3.5 in real-world scenarios.
“With Llama 3, we set out to build the best open models that are on par with the best proprietary models available today,” said Meta in a blog post announcing the release.
The initial Llama 3 models being opened up are 8 billion and 70 billion parameter versions. Meta says its teams are still training larger 400 billion+ parameter models which will be released over the coming months, alongside research papers detailing the work.
Llama 3 has been over two years in the making with significant resources dedicated to assembling high-quality training data, scaling up distributed training, optimising the model architecture, and innovative approaches to instruction fine-tuning.
Meta’s 70 billion parameter instruction fine-tuned model outperformed GPT-3.5, Claude, and other LLMs of comparable scale in human evaluations across 12 key usage scenarios like coding, reasoning, and creative writing. The company’s 8 billion parameter pretrained model also sets new benchmarks on popular LLM evaluation tasks:
“We believe these are the best open source models of their class, period,” stated Meta.
The tech giant is releasing the models via an “open by default” approach to further an open ecosystem around AI development. Llama 3 will be available across all major cloud providers, model hosts, hardware manufacturers, and AI platforms.
Victor Botev, CTO and co-founder of Iris.ai, said: “With the global shift towards AI regulation, the launch of Meta’s Llama 3 model is notable. By embracing transparency through open-sourcing, Meta aligns with the growing emphasis on responsible AI practices and ethical development.
”Moreover, this grants the opportunity for wider community education as open models facilitate insights into development and the ability to scrutinise various approaches, with this transparency feeding back into the drafting and enforcement of regulation.”
Accompanying Meta’s latest models is an updated suite of AI safety tools, including the second iterations of Llama Guard for classifying risks and CyberSec Eval for assessing potential misuse. A new component called Code Shield has also been introduced to filter insecure code suggestions at inference time.
“However, it’s important to maintain perspective – a model simply being open-source does not automatically equate to ethical AI,” Botev continued. “Addressing AI’s challenges requires a comprehensive approach to tackling issues like data privacy, algorithmic bias, and societal impacts – all key focuses of emerging AI regulations worldwide.
”While open initiatives like Llama 3 promote scrutiny and collaboration, their true impact hinges on a holistic approach to AI governance compliance and embedding ethics into AI systems’ lifecycles. Meta’s continuing efforts with the Llama model is a step in the right direction, but ethical AI demands sustained commitment from all stakeholders.”
Meta says it has adopted a “system-level approach” to responsible AI development and deployment with Llama 3. While the models have undergone extensive safety testing, the company emphasises that developers should implement their own input/output filtering in line with their application’s requirements.
The company’s end-user product integrating Llama 3 is Meta AI, which Meta claims is now the world’s leading AI assistant thanks to the new models. Users can access Meta AI via Facebook, Instagram, WhatsApp, Messenger and the web for productivity, learning, creativity, and general queries.
Multimodal versions of Meta AI integrating vision capabilities are on the way, with an early preview coming to Meta’s Ray-Ban smart glasses.
Despite the considerable achievements of Llama 3, some in the AI field have expressed scepticism over Meta’s motivation being an open approach “for the good of society.”
However, just a day after Mistral AI set a new benchmark for open source models with Mixtral 8x22B, Meta’s release does once again raise the bar for openly-available LLMs.