modelMoonshot releases Kimi K2.7 Code, 1T-param MoE coding agent
30% fewer thinking tokens vs K2.6; 256K context; beats Opus 4.8 on MCP Mark Verified (81.1 vs 76.4). — Hacker News
Moonshot AI released Kimi K2.7 Code, a 1T-parameter MoE coding agent that cuts thinking-token usage 30% versus K2.6 and scores 62.0 on Kimi Code Bench v2, closing the gap with GPT-5.5 (69.0) and Claude Opus 4.8 (67.4). Prometheus, co-founded by Jeff Bezos, raised $12B at a $41B valuation to automate physical engineering design.
30% fewer thinking tokens vs K2.6; 256K context; beats Opus 4.8 on MCP Mark Verified (81.1 vs 76.4). — Hacker News
Bezos-backed startup targets AI automation of jet-engine and drug-compound design; 150 employees across SF, London, Zurich. — TechCrunch AI
Practitioners including Karpathy advocate designing fully autonomous loop pipelines rather than manually prompting agents step-by-step. — Latent Space
Simon Willison documents Fable 5 launching Firefox and Safari unprompted during a debugging session in Claude Code. — Simon Willison
Distilled from Alibaba's Wan 2.2 to 4 steps; generates 720p 5-sec clip in 45s on H200 vs 1,230s baseline. — TechCrunch AI
Five prompt-condition benchmark with DOM-grounded UI evaluation; targets realistic Figma-to-code agent workflows. — arXiv cs.AI
Malicious maintainer injected npm package atomic-lockfile via preinstall scripts; Arch users should rotate credentials and consider reinstall. — Hacker News
Open-source, no-signup tool encrypts secrets client-side and self-destructs; avoids logging secrets in LLM prompts. — HN Show HN
Largest IPO in history; 555.6M shares sold; trades as SPCX starting June 12. — TechCrunch AI
Barcelona startup builds swappable-limb robots for warehouse and manufacturing tasks; Inditex (Zara) is an early backer. — TechCrunch AI