In April 2024, Meta released Llama 3, the next generation of its open source language model family. The released variants included an eight billion parameter model and a seventy billion parameter model, with a four hundred billion parameter version under continued training. The benchmarks showed substantial improvements over Llama 2, with the seventy billion parameter model approaching GPT-3.5 capability on many tasks and showing strength in coding and reasoning that previous open source models had struggled with.
The licence remained permissive enough for commercial use under the same scale threshold that Llama 2 had used. The combination of strong capability and a usable licence meant that Llama 3 was immediately the most attractive option for organisations that wanted to run their own language model infrastructure rather than depending on a proprietary API.
The release confirmed that Meta was serious about its open source AI strategy in a way that had been merely plausible after Llama 2. The investment required to train models at this scale and release them openly was enormous. Meta was making that investment as a recurring commitment rather than a one-time experiment. The strategic logic remained as before. Open releases damaged the moats of closed-model competitors, and positioned Meta at the centre of the open AI ecosystem.
The fine-tuning ecosystem responded immediately. Specialised variants of Llama 3 for code generation, for various non-English languages, for medical and legal applications, for safety filtering, and for many other purposes appeared within weeks. The serving infrastructure improvements that had come with Llama 2 continued to mature. The cost of running Llama 3 inference dropped quickly as more providers competed to host it efficiently.
What the release also did was force a recalibration of the closed-versus-open conversation around AI. Through 2023, the argument had been that proprietary models would maintain a meaningful capability lead and that open models would always lag. Llama 3 made that argument harder to maintain. The leading proprietary frontier models were still ahead of the leading open models. The gap had narrowed to a degree that meant for many practical applications, the choice between open and closed would be made on the basis of cost, control, and integration rather than capability alone.
The four hundred billion parameter model that was still training would be released later in 2024 and would push the open frontier closer to the proprietary frontier than many observers had expected to be possible at this point.