Top AI News Weekly
GLM-5.1 tops SWE-Bench Pro as the first open-source model beating GPT-5.4 and Claude Opus. Meta launches Muse Spark multimodal reasoning model with multi-agent orchestration. Anthropic announces Project Glasswing for AI-powered cybersecurity. LPM 1.0 achieves first full-duplex conversational video generation. Waypoint 1.5 and InSpatio-World advance real-time 3D worlds. ACE-Step 1.5 brings commercial-grade music generation to consumer GPUs.
23 launches and research drops that matter for enterprise AI builders—curated, tagged, and ready for your next roadmap sync.
New drops
23
Unique sources
23
Key themes
video · 3d · image
New reasoning systems, world models, and alignment papers.
Open-source 744B MoE flagship model that tops SWE-Bench Pro at 58.4, beating GPT-5.4 and Claude Opus 4.6. Designed for agentic engineering with 8+ hour autonomous task execution, released under MIT license.
View release ↗Natively multimodal reasoning model with tool-use, visual chain of thought, and multi-agent orchestration via Contemplating Mode that enables parallel reasoning across multiple agents.
View release ↗Embodied agents learning to act in complex worlds.
Open-source AI memory system that stores conversations locally in a structured palace architecture (wings, rooms, halls), achieving 96.6% recall on LongMemEval benchmarks without cloud services.
View repo ↗Python MCP server and CLI for programmatic Google Flights search, enabling AI assistants like Claude to search flights through natural language conversation.
View repo ↗Open-source local-first AI assistant that builds a knowledge graph from email and meeting notes, then uses that context for drafting, meeting prep, and document generation with BYO LLM support.
View repo ↗Personal knowledge management system that creates a searchable AI-augmented brain from meetings, emails, and notes, enabling AI agents to maintain long-term memory via markdown-based knowledge base.
View repo ↗Frameworks, playbooks, and OSS repos.
KV cache compression for LLMs using block-diagonal rotations instead of full-rank transforms, achieving better perplexity, faster decode speeds, and 44x fewer parameters than Google's TurboQuant.
View repo ↗MCP-integrated tool that creates structural codebase maps using Tree-sitter, providing precise context to AI coding assistants and reducing token consumption by 8.2x during code review.
View repo ↗Comprehensive API platform providing unified access to numerous AI models for generating and editing images, videos, and audio content through a single interface.
View release ↗