Meta Llama 4 One Year On: What Builders Actually Ship With

Meta released the Llama 4 family in April 2025 with three models named for different scale-and-speed trade-offs: Scout for small and fast, Maverick for the production sweet spot, and Behemoth for frontier-scale tasks. Thirteen months on, the Llama 4 family has become the most-deployed open-source LLM lineage in production, with adoption that has reshaped the cost curve for AI features inside vertical SaaS and white-label app platforms.

Menlo Park · May 17, 2026

What happened

Meta announced Llama 4 on April 5, 2025, with the immediate release of Scout (17 billion active parameters, 109 billion total in a mixture-of-experts setup) and Maverick (17 billion active, 400 billion total). Behemoth, the largest model in the family, was previewed but kept in training. Meta’s official AI blog introduced the family with benchmark numbers and the architectural details — most notably the move to a mixture-of-experts design that materially changed the cost of running these models compared with their Llama 3 predecessors.

The Llama 4 release also marked a shift in Meta’s open-source posture. The license was tightened versus Llama 3 — companies with over 700 million monthly active users now need an explicit commercial agreement — but the practical effect on smaller builders and white-label shops was nil; the threshold is high enough that the typical developer is unaffected. The Verge covered the launch and the license change in detail. In the 13 months since, Llama 4 Maverick has become the default open-source model running behind the scenes of countless vertical SaaS products, white-label apps, and AI-features-as-a-service platforms.

Why it matters for builders and founders

If you ship any product that uses AI for high-volume background tasks — content moderation, customer-message classification, document parsing, summarisation, lead enrichment — Llama 4 Maverick is probably the cheapest credible option that gets the job done. Running Maverick on a managed inference provider like Together AI, Fireworks, or Groq costs a fraction of what the equivalent volume on Claude or GPT-5 would cost, and for most non-frontier workloads the quality difference is small enough to be invisible to end users.

For white-label app shops specifically, this has compressed margins on AI-feature pricing. Two years ago, a SaaS shipping AI features had to charge a meaningful premium because the inference bill was real. In 2026, the inference bill on Maverick-class workloads is small enough that AI features are increasingly bundled into base subscriptions rather than gated behind an AI tier. The competitive dynamic forced this: any vendor that tried to price AI as a separate $20-a-month add-on lost customers to whoever bundled it.

The details, in plain English

“Mixture of experts” — abbreviated MoE — is an architectural technique where a model contains many specialised sub-networks (the “experts”), but only a small subset of them are activated for any given input. The advantage is that the model can have a very large total parameter count (which improves capability) while running with the inference cost of a much smaller model (which improves economics).

For Llama 4, this means:

• Scout — 109 billion total parameters, 17 billion active per token. Optimised for fast, cheap inference. The model most teams reach for when latency matters and the task is not pushing capability.
• Maverick — 400 billion total, 17 billion active per token. The production sweet spot. Quality competitive with the smaller frontier models, at a fraction of the cost.
• Behemoth — over 2 trillion total parameters. Still being released in stages through 2025 and into 2026. Designed to compete on raw capability with Claude Opus and GPT-5, with the licensing constraints of the larger Llama family.
• Native multimodality — Llama 4 was the first Llama generation trained natively on text and image inputs together, rather than bolted on after the fact. The result is materially better at image understanding than Llama 3 was.

The practical reason MoE matters: a 17-billion-active-parameter model is something you can serve on a single GPU per replica, which means inference cost scales linearly with traffic in a way that frontier dense models do not.

The bigger picture

The success of Llama 4 has done two important things for the AI ecosystem. First, it has created a credible open-source second tier that keeps the frontier labs honest on pricing. Anthropic and OpenAI cannot raise per-token prices arbitrarily because any meaningful price increase pushes more workloads to Maverick or DeepSeek. Second, it has enabled an inference-provider ecosystem — Together AI, Fireworks, Groq, Cerebras, and others — that competes on serving the same open-source models at different latency and price points. The result is a market for inference that did not really exist in 2023.

Meta’s motivation for shipping Llama as open-source remains contested. The most charitable reading is that Meta benefits when AI infrastructure is commoditised because Meta is a consumer-product company whose competitive advantage is distribution, not model quality. The less charitable reading is that Meta needs to be a serious player in AI infrastructure to attract and retain the talent that powers its consumer products, and open-source releases are how it signals that. Either way, the practical outcome is the same: builders get a free, capable model family with permissive enough licensing.

What to watch next

Three things to watch through the rest of 2026. First, the eventual full release of Behemoth and whether it lands close to frontier capability or trails by a wider margin — if it lands competitively, the cost calculus for frontier workloads shifts again. Second, whether Meta keeps the Llama license stable through Llama 5; the 700 million MAU threshold has not been tested in court yet, and any change to it would affect how startups can grow without renegotiating. Third, the inference-provider consolidation: with 5-plus serious providers competing on Maverick, some will not survive, and the question of which two or three become the default infrastructure layer matters more than the model release itself.

For builders today, the practical move is to evaluate Maverick on your highest-volume AI workload — whatever you currently send to a frontier API. If the quality is acceptable, switching saves real money. If it is not, the evaluation is still useful because it tells you which workloads genuinely need frontier capability and which were over-buying.