Meta Releases LLaMA 4: Open Weights, Frontier Performance
Artificial Intelligence
Covers AI systems, large language models, and their real-world implications.
Technology
Reports on software, hardware, and the companies shaping the digital world.
Finance & Markets
Analysis of markets, economic indicators, and financial developments.
Meta has released LLaMA 4, its most capable open-weights model to date, and the results have shaken assumptions about the performance ceiling of openly available systems. On the MMLU, HumanEval, and MATH benchmarks, LLaMA 4 scores within a few percentage points of GPT-4o — a gap that would have seemed impossible two model generations ago.
The release comes with full weights available for download under a permissive research licence, meaning developers, academics, and smaller companies can run the model on their own infrastructure without API costs or usage restrictions. For applications involving sensitive data — healthcare, legal, finance — this is a meaningful shift. Running a frontier-class model behind your own firewall is no longer a fantasy reserved for the largest enterprises.
Meta has framed the release explicitly as a strategic counter to closed ecosystems. The company argues that open models drive faster innovation, reduce single-point-of-failure risks in AI supply chains, and allow safety researchers to study model internals rather than probing them from the outside.
Fine-tuning results published alongside the release show that LLaMA 4 can be adapted to domain-specific tasks in hours on consumer-grade hardware, a workflow that was impractical with earlier open models at this capability level.