Meta Releases Llama 4: Scout and Maverick Debut as Natively Multimodal Open Models

Meta released the first two Llama 4 models on April 5, 2025: Llama 4 Scout and Llama 4 Maverick — the first natively multimodal, open-weight Llama models, both using MoE architectures with 17B active parameters.

Scout's signature feature is an industry-leading context window of up to 10 million tokens, while Maverick targets the performance/cost balance; the larger Behemoth was still in training at the time.

Polarized Reception: The Benchmark-vs-Feel Gap

Community testing split sharply after launch: some developers felt real-world performance lagged the official numbers, and controversy over an arena build differing from the released weights compounded the trust erosion. For an open-source flagship, reputation is the moat — the crack loosened the 'default to Llama' instinct, and developers voted with their feet toward Qwen and DeepSeek.

Behemoth's Stall and the Aftershocks

The flagship Behemoth slipped repeatedly and never shipped as planned, becoming the emblem of Meta's AI setback. Two months later Zuckerberg poured $14.3 billion into Scale AI, stood up Superintelligence Labs and launched the nine-figure hiring spree (see our coverage) — the direct fuse of that gamble was the anxiety of Llama 4 failing to hold the open-source lead.

The Open-Source Throne Changes Hands

Llama once single-handedly defined the open-model ecosystem, but 2025's open-source story was rewritten by the Chinese relay: DeepSeek R1 detonated open reasoning, Qwen3 covered every size class, Kimi K2 topped the scale charts (see our coverage). With OpenAI returning to the table via gpt-oss, Llama went from synonym-for-open-source to one player among several.

Our Take

Llama 4's sample lesson is the brutal logic of open ecosystems: leaders hold no tenure, and once trust slips, developer migration costs approach zero. It also explains Meta's radical pivot that followed — buying talent, rebuilding, and reported closed-source deliberations: when openness stops buying ecosystem position, openness itself gets re-evaluated.