HyperFlow H3 Flash: local video-with-audio diffusion in 6 steps
A 14B INT8 cut of MiniMax-H3 runs open in ComfyUI; Cursor adds live-app publishing and voice agents, and a cheap LLM judge hits Opus accuracy for ~4x less.

Copy markdown
Open video, with sound, in six steps
HyperFlow H3 Flash compresses MiniMax-H3 (~31B) into a 13.7B, 4-8 step model with INT8 and GGUF builds, so you can generate joint video+audio locally through its ComfyUI nodes. It's Apache-2.0, which turns synced text-to-video-with-audio into a self-hostable weekend project instead of a metered API call.
Cursor ships deployable apps and voice agents
Cursor's cloud agent can now build an app and publish it to a live domain, with Postgres, SSO auth, hosting and secret-scanning included, on any paid plan with cloud agents. A day later it added real-time voice calls with agents across the IDE, web and iOS ("Hey Siri, call Cursor"), running on Grok Voice with zero-retention Privacy Mode.
A cheap judge that phones Opus only when unsure
Hightouch's engineering post (surging on Hacker News) lays out a hybrid LLM-judge harness: a fast, cheap decision model scores each case and escalates only the uncertain 30-90% confidence band (~22% of items) to Claude Opus 5.5. On their identity-resolution benchmark that hit 98.8% accuracy at $0.88 per 1,000 vs Opus-alone's 99.0% at $3.86 - roughly 4x cheaper, and a pattern worth copying for any LLM-as-judge or eval pipeline.