@FastAPI surpassed Flask in weekly downloads this past week

💻 NVIDIA introduces Universal Deep Research, a system that lets users plug in any LLM and run custom research strategies as code. This breaks from fixed deep research products by handing over control of model choice, source rules, verification, and spend. A natural language strategy is compiled into 1 generator function that yields structured progress updates while calling search and the model. Intermediate facts live in named variables rather than the prompt, so full runs fit in about 8K tokens and remain traceable. Tools are called synchronously for deterministic behavior, and the model acts as a local helper for summarizing, ranking, or extracting. Control flow runs on CPU while targeted model calls touch small text slices, which lowers GPU usage, latency, and cost. Reliability improves because generated code mirrors each step with explicit comments, which reduces skipped steps and surprise constraints. Execution is sandboxed to block host access, keeping strategies confined to a safe runtime. The demo UI lets users edit strategies and use minimal, expansive, and intensive templates from single pass to iterative loops. Limits include dependence on the model's code generation, the quality of the user's plan, and little mid run steering beyond stop. For teams this means bring your own model and method, so best models pair with domain playbooks without vendor lock. The view here is that the code driven control with scoped model calls makes deep research fast, cheap, and auditable. A rich strategy library will drive adoption, because most users will not author complex plans. ---- Paper – arxiv. org/abs/2509.00244 Paper Title: "Universal Deep Research: Bring Your Own Model and Strategy"