Meta released Llama 4 in April 2025. I’ve been using and watching it for about a year now, which feels like a good vantage point to say something useful about what it actually meant — not what the launch blog post said it would mean, but what happened.
The launch was a big deal. Meta released two models: Llama 4 Scout and Llama 4 Maverick. Both are multimodal — they handle images and text. Both use a mixture-of-experts architecture, which means the total parameter count is large but only a fraction is active during inference. Scout has 17 billion active parameters out of 109 billion total. Maverick has 17 billion active out of 400 billion total. The benchmark numbers at launch put them competitive with GPT-4o and Gemini on several tasks. Meta also claimed Scout had a 10 million token context window, which was a striking number.
What the Context Window Claim Actually Means
I want to spend a moment on this because “10 million token context” is the kind of specification that sounds transformative and needs some unpacking.
Yes, the model was trained with a very long context window. Whether you can actually use 10 million tokens effectively — whether the model pays useful attention to information buried deep in an extremely long context — is a separate question from whether it technically accepts that input. Research on long-context LLMs consistently shows that performance degrades as the relevant information gets farther from the ends of the context. “Lost in the middle” is a documented problem with current architectures.
In practice, long context is very useful. Being able to hand a model a 200-page document, or an entire codebase, and ask questions about it — that’s real capability that Llama 4 Scout genuinely has. Whether every token of a 10M token input is being effectively attended to is a harder question with a less clean answer.
The Open Weights Bet
The more consequential thing about Llama 4 isn’t the benchmark scores — it’s that Meta continues to bet heavily on open weights.
This matters for a few reasons. First, it creates a competitive floor that keeps closed model pricing honest. If a capable open model is freely available, the closed models need to justify their pricing with meaningful capability advantages. Second, it puts powerful models in the hands of organizations with compliance or privacy constraints that prevent using external APIs. Third, it enables fine-tuning and customization that you simply can’t do with closed models.
A year on from the launch, Llama 4 has been fine-tuned, quantized, and run in more configurations than Meta probably anticipated. The community adoption has been strong. Running a quantized version of Scout locally is feasible on high-end consumer hardware; Maverick requires serious infrastructure, but it’s accessible via hosted inference from multiple providers.
Where Llama 4 Actually Stands Now
Honest assessment: the base models are excellent. For coding tasks, particularly Python, JavaScript, and general-purpose scripting, Scout is competitive with frontier closed models on tasks that fit reasonably within context. The multimodal capability is real — image understanding works, and it’s useful. The instruction-following quality improved meaningfully from Llama 3.
Where the gap with frontier closed models shows up: complex multi-step reasoning, tasks requiring very precise instruction following on novel or unusual formats, nuanced writing that requires sustained coherence over long outputs. These aren’t fatal weaknesses — they’re places where you notice you’re using a different class of model.
Fine-tuned Llama 4 variants have produced some interesting specialized results. Organizations that can invest in fine-tuning for their specific domain get something the base model doesn’t offer. That’s a real advantage of the open-weight approach.
What Meta’s Bet Means for the Broader Market
A year out from the Llama 4 launch, a few things seem clearer than they did in April 2025.
The “open vs closed” framing is getting less useful. The more interesting question is: what can you accomplish with what tradeoffs? Llama 4 is genuinely capable enough that for a large number of use cases, the tradeoff calculus shifts toward open — not because it’s free (you still pay for compute) but because the capability gap has narrowed to where the control and flexibility advantages can dominate.
Meta’s strategy appears to be: make the foundation free, create value through the ecosystem. If the best open models are Meta models, Meta’s infrastructure, cloud services, and developer tools are well-positioned. It’s not purely altruistic, and that’s fine — the incentives happen to produce something useful for everyone else.
The mixture-of-experts architecture that Llama 4 uses has also become more widely adopted. When you can get a model that performs like a 400B parameter model but only activates 17B parameters at inference time, the cost/capability ratio looks very different from dense models at the same capability level. This wasn’t new with Llama 4 (Mixtral did it earlier), but Meta’s adoption of it at scale normalized the approach.
A Year In
Llama 4 turned out to matter. Not because it immediately matched the frontier closed models — it didn’t, and doesn’t — but because it raised the capability floor for the open ecosystem high enough that the open-weight choice became defensible for a much wider set of use cases. That’s a different kind of impact than beating a benchmark, and I think it’s the more durable one.
Meta will release Llama 5, presumably. The pattern they’ve established — competitive open weights released on a reasonable cadence — is more valuable to the field than any single model release. A year out from Llama 4, that pattern looks well-established.