Hugging Face AI Updates: April 22, 2026

1. TII Launches QIMMA, an Arabic LLM Leaderboard That Validates Benchmarks Before Ranking Models

Hugging Face. Technology Innovation Institute published QIMMA (قِمّة, “summit”), an Arabic LLM leaderboard consolidating 109 subsets from 14 source benchmarks into more than 52,000 samples across seven domains. The distinguishing move is a quality-validation pipeline that filters the input benchmarks themselves before any model is evaluated, aimed at stopping scores from being inflated by noisy or mis-translated items. Across 46 evaluated models, Qwen3.5-397B-A17B-FP8 tops the table at 68.06, followed by AIC’s Karnak (66.20) and Inception’s Jais-2-70B-Chat (65.81). Source

2. Hugging Face Argues Openness Is a Structural Advantage in AI-Era Cybersecurity

Hugging Face. Margaret Mitchell, Yacine Jernite, Clem Delangue and collaborators published a policy piece arguing that as cybersecurity capability gets bundled into LLM-powered systems like Anthropic’s Mythos, open-source components shift the attacker-defender balance in favor of defenders. The thesis: open ecosystems distribute vulnerability detection, verification, coordination, and patch propagation across many hands, while closed stacks create single points of failure and make sensitive data flow through external providers. The authors explicitly advocate semi-autonomous defensive agents with prespecified actions, human approval steps, auditable logs, and deployment inside the organization’s own infrastructure. Source