Researchers at Epoch AI demonstrated that the widely available open-weight model GLM 5.3 autonomously converted public Firefox security fixes into working exploits affecting six Firefox releases, each costing approximately $1,000 in API fees to generate and bypassing safeguards from multiple tested providers (The Decoder). The exploits were produced without human input in under 48 hours, illustrating the capability gap between open-weight and frontier models on security-critical tasks.
Mistral released weights for Large 4, also known as “Le Chonk,” a trillion-parameter model trained entirely in Europe (The Decoder). The model achieved strong performance on coding benchmarks relative to other open-weight alternatives and operates at competitive inference cost, though independent testing shows it trails leading Chinese models on overall capability metrics (TechCrunch). Separately, Baseten released joey-1, a 2B open-weight decision model that returns calibrated probabilities at $0.01 per million input tokens with 20 ms P50 latency (MarkTechPost), and Prime Intellect released Prime Mega MoE, a trainable fused MoE kernel achieving up to 3.11x forward pass speedup and 34% training step improvements on Qwen3 MoE (MarkTechPost).