Meta's ambitious Llama 4 Behemoth, a 2-trillion parameter AI model built on a Mixture-of-Experts architecture, remains 'in training' a full year after its expected release. Unveiled in April 2025 with plans to serve as a foundational 'teacher model' for open-weight AI, the model has missed its launch window entirely, with no new delivery date announced.
The delay has proven costly. In the past year, competitors including Gemini 2.5 Pro, Claude Opus 4.7, GPT-5.4, and Kimi K2.6 have shipped and set new performance benchmarks, reshaping the competitive landscape. The prolonged timeline raises doubts about whether Behemoth's original design can still compete or if market dynamics have made it obsolete before release.
For Meta, the setback underscores the challenges of scaling cutting-edge AI and the risks of an aggressive roadmap. The company has not commented on revised timelines or whether Behemoth's architecture will be adapted to catch up with newer models.