Cloudflare Blog, summarized in your digest

Cloudflare released Clef-omni, an open-weight decision model that processes text, images, audio, and video in one API call, using a Qwen3-Omni-30B-A3B-Instruct foundation with schema-constrained scoring. It delivers median latency of about 130 ms for text and 1.5 seconds for a 21-second video. Clef-flash now costs $0.038 per million input tokens, but its hosted context window is 24k; Clef latency improved 1.7–2×, while Clef-omni costs $0.15 per million tokens.

Cloudflare introduced on-demand CPU and memory profiling for Workers and Durable Objects, available through the Workers Observability dashboard or CLI. Developers can profile active production isolates, inspect interactive flamegraphs, and download profiles. Flamegraphs identify resource-intensive functions: optimizing recursive JSON processing made `genericR2JsonReplacer` 2.7x faster, while removing partially disabled Prometheus instrumentation reduced P999 memory from 133 MB to 118 MB. Profiling requires sufficient traffic; Cloudflare is developing continuous profiling for rare events.

The Deno team is joining Cloudflare to simplify self-hosting Workers and Durable Objects, allowing developers to use the same primitives in more environments. Ryan Dahl and Kenton Varda explain the motivation and background behind the move, which connects Deno’s work with Cloudflare’s developer platform.

Cloudflare redesigned Radar to make its real-time Internet data more accessible without sacrificing technical depth. The new landing page uses a map to foreground traffic and outages, then guides users through country insights, trends, and data stories via a tabbed layout. Its design combines Cloudflare’s marketing, content, and product patterns and migrates to Kumo components. The redesign supports Radar’s broader goal of delivering actionable security, performance, and traffic insights to a wider audience.

Cloudflare built an evidence-grounded, multi-AI-agent security operations harness for Managed Defense that triages alerts, gathers deterministic reconnaissance, correlates related events, and produces auditable recommendations while keeping analysts responsible. Clef filters noise and scores evidence; specialist agents separately analyze traffic, customer history, global aggregate telemetry, and threat intelligence, then a synthesis agent cites a versioned evidence package. Workers, Workflows, D1, R2, Durable Objects, Flue, and AI Search support the process, which records gaps and avoids unsupported classifications.