Meta has launched a new version of its Large Language Model (LLM) called Llama 3.3. This model, with 70 billion parameters, is designed to be simpler and more cost-effective to operate. Ahmad Al-Dahle, Vice President for Generative AI at Meta, announced Llama 3.3 on social media, comparing it with other models like Amazon’s Nova Pro, Google’s Gemini Pro 1.5, and OpenAI’s ChatGPT-4o.
Meta has shared various details about its Llama model family, though specific training data remains undisclosed. Llama 3.3 performed well in several benchmarks, notably in “Instruction Following,” where it excelled in following given prompts accurately. The model was tested using the IFEval benchmark, which includes around 500 prompts with verifiable tasks. Llama 3.3 successfully answered 92.1% of these prompts, matching the performance of Amazon Nova Pro.
The model also excelled in “Long Context” prompts, achieving a 97.5% success rate. This was slightly lower than the older Llama model 3.1, which scored 98.1%. This test involved identifying specific sequences in large datasets, akin to finding a needle in a haystack. Llama 3.3 also showed strong performance on the Multilingual MGSM dataset, solving 250 school-level math problems in ten languages with a 91.1% success rate. Though slightly less than Llama 3.1’s 91.6%, the newer model offers operational and cost-efficiency improvements.
Meta CEO Mark Zuckerberg announced that Llama models have been downloaded 650 million times, with 600 million people using them monthly. Llama models are available for research and commercial use under certain conditions. Platforms with over 700 million monthly active users require a special license from Meta.
In November, it was revealed that the Chinese military uses Meta’s Llama models, prompting Meta to allow the US government to use its AI for national security purposes. Recently, Meta decided not to release a Llama version in the EU due to regulatory concerns.
Looking ahead, Meta plans to release Llama 4, which is expected to require ten times the computational power for training compared to its predecessors. The release of Llama 4 is anticipated in 2025.