Together AI released Together Compiler, an agent-driven optimization loop that monitors inference traffic and re-compiles the forward pass in real time to match shifting patterns (MarkTechPost). In testing across multiple models, the system outperformed the best existing InferenceX configurations by up to 2.6x tokens per second within 24 hours per model, addressing a core cost vector for inference-heavy deployments.
Box announced an agent-first content retrieval system powered by Nvidia’s Nemotron, claiming four times faster query response, 20% fewer errors, and 90% fewer tokens consumed compared to prior approaches (TechCrunch). Separately, Ricursive Intelligence released Ripple-65B Preview, a 65B open-weight dense model scoring 80.5 on Terminal-Bench 2.1 and outperforming much larger open models (MarkTechPost), and webAI released TwIL-8B, an 8.8B reasoning model post-trained from IBM Granite that scored 98.2 on MATH-500 (MarkTechPost).