Which AI writes the better take? You decide — blind.

975B-parameter open-weights model with only 41B active at inference, cutting GPU memory needs by ~20× while matching top-tier performance on code, math, and multimodal tasks. This lets you deploy a single model that handles text, images, and speech in production without swapping architectures, and fine-tune it on your own data for domain-specific workflows without hitting memory walls. Expect lower cloud costs and faster iteration, but watch for calibration drift if you push thinking-time too low.

Record-level provenance cuts over-deletion from 101× to 1.3× when honoring data-removal requests. This means you can now comply with revocation demands without nuking entire datasets or retraining from scratch, saving weeks of compute and preserving model performance—just add 1–4% pipeline overhead.

Robotics novices using SPINE, an agentic AI framework, can now deploy bimanual robots with 100% success rate, compared to 75% without it, and reduce mean time-to-teleoperation by nearly 3 minutes; this shift enables faster and more reliable deployment of embodied AI in production environments, reducing dependence on expert calibration.

An automated red teaming system achieved a 2x improvement in detecting and mitigating prompt injection attacks through self-play. This capability directly enhances the robustness of production LLMs, reducing the manual effort and cost associated with vulnerability testing and patching. This enables teams shipping LLM-based applications to more reliably prevent exploits and improve overall security posture.

Cache hit rates can cut effective input costs dramatically, with one model's lower cache-read pricing resulting in nearly half the total cost of another despite higher base pricing. This means model routers must account for serving infrastructure and workload patterns, not just model pricing, to optimize costs for agent workloads that reuse context across steps.