Meta Llama 4 Reshapes Open Source AI Landscape

Honestly, Meta really “dropped the bomb” this time. Llama 4 came out with three variants - Scout, Maverick, and the flagship Opus - and the top performer hit 92 on MMLU, surpassing every open-source model I’ve tested and even some closed-source ones.

What surprised me most was the context window. Llama 4 Opus supports 2 million tokens - that’s equivalent to reading the entire “Three-Body Problem” trilogy with room to spare. Open-source models have always been weak on long contexts, but Meta just shattered the ceiling this time.

Of course, what I care about most is actual deployment cost. Maverick has only 17B parameters but performs close to GPT-4o, which means running top-tier models on consumer GPUs isn’t a dream anymore. I’m personally excited to try running one locally on a 4090 - that’s the real point of open source: making it accessible to everyone.

But what intrigues me most is Meta’s licensing adjustment. Llama 3’s “no commercial use” clause got a lot of backlash. How will Meta adjust this time? I’d expect one truly open-source version and one more commercial-friendly version. Commercial companies don’t want their core competitiveness built on legally risky foundations.

Overall, Meta’s strategy is clear: use open-source models to force closed-source vendors to cut prices while attracting the developer ecosystem. My personal take is that once open source reaches scale, the premium pricing of closed-source models will be severely squeezed. For people like us working with large models daily, this is genuinely good news.