MiniMax-H3: open-weight video with native audio, up to 2K, 33B
The weights make 4–15s clips with 32kHz stereo sound; LTX-2.5 pushes open video to 4K, and Cursor now watches every deploy per-environment.

Copy markdown
Open weights, video with sound, up to 2K
MiniMax-H3 is a 33B dense transformer that turns text, images, or video into 4–15s clips at 24fps with native 32kHz stereo audio, up to 2K and six aspect ratios across 11 languages. The weights ship under a community license, so you can self-host or fine-tune a video-and-sound generator instead of renting one per second.
LTX-2.5 takes open video to 4K — free under $10M
Lightricks' LTX-2.5 does text/image/video-to-video with synced audio up to 4K (3840×2176) and ships an 8-step distilled variant for speed. Commercial use is free for anyone under $10M revenue — a realistic base for shipping a video-gen product this weekend.
Cursor bolts a Rollouts bot onto every PR
Cursor's new Rollouts bot monitors each pull request through deploy, reports change health per environment, and can open a revert PR or hand a fix to a cloud agent on a regression. A companion Security Review bot flags SQL/command injection, auth bypasses, and committed secrets with severities and fixes; both come with free credits (~50 Teams, ~500 Enterprise) for the first 10 days.
27B-class quality in 5.95GB, at 1.72 bits
Prism ML's Ternary-Bonsai-2-27B is a ternary GGUF quant of Qwen3.8-27B (Apache 2.0) claiming 98.2% of FP16 intelligence at 1.72 bits per weight — shrinking a 54GB model to ~5.95GB and running ~47 tok/s on an Apple M5 Max, ~130 on an RTX 5090. Frontier-ish local inference now fits a laptop.
MiniCPM5-2B for on-device, 131K context
OpenBMB's MiniCPM5-2B is a 2B-class model with a 131,072-token context window and tool-calling, Apache 2.0, tuned for phones and edge boxes — an API-free LLM you can wire straight into a local app.