By olud.ai editorial · built from our own tracking data, published daily
Langfuse v4.11.0 shipped with a fix that makes observation data honest when it exceeds the load cap, tightening transparency for teams monitoring LLM traces. The release is one of eight minor updates across the ecosystem today, including n8n, Ollama, MinerU, Cline, CrewAI, Pydantic AI, and llama.cpp.
The open-source AI ecosystem gained 270,384 GitHub stars today across the tracked project index, with 86 new projects added. The weekly total for new projects stands at 1,921, alongside 100 new Hugging Face spaces. The top gainer today is "skills" with 7,887 stars.
Releases of the day
Langfuse v4.11.0 addresses trace honesty past load caps, a change that affects users monitoring high-volume pipelines. n8n 2.34.6 fixes the details field in continueErrorOutput mode and migrates Google Ads from API v21 to v25. Ollama v0.32.11 adds support for DeepSeek Harness in its launch command and integrates Meta's Muse Code agentic coding CLI. MinerU 3.4.5 fixes docx tables that were silently dropped when cells contained non-text special characters. Cline v3.0.55 changes auto-update behavior so installations wait while a CLI is attached to the Hub. CrewAI 1.15.16 introduces execution context management with UUID support and records flow-ending exceptions. Pydantic AI v1.107.5 closes a security advisory for DNS rebinding in its local dev web chat UI. llama.cpp b10427 optimizes q4_K dense FFN on Intel GPUs by fusing gate, up, and GLU multiplications.
Input token prices moved across seven models today. Google cut Gemini 3.6 Flash input price by 50 percent to $0.75 per 1M tokens, with its batch variant down 49 percent to $0.38. Z.AI reduced GLM 5.2 input by 20 percent to $0.39 and GLM 4.6 input by 9 percent to $0.50. Moonshot AI lowered Kimi K2.6 input by 3 percent to $0.56. Alibaba raised Qwen3 Next 80B A3B Instruct input by 11 percent to $0.10 and Qwen3.5-35B-A3B input by 79 percent to $0.25.
A Hugging Face community paper by Yuanyang Yin, Gongxuan Wang, Yifan Zhan and others presents Alaya-EVOKE, a method for interactive world models that avoids holding long history in the denoiser context or key-value cache. The approach addresses conflicting demands between persistent memory, responsive interaction, and long-horizon generation.
Today's releases lean toward infrastructure polish and security fixes, while the price drops from Google and Z.AI continue a downward trend in inference cost. The sharp price increase for Qwen3.5-35B-A3B stands out as an exception.
Source: olud.ai tracking of 10,000+ open-source AI projects, 300+ models and live provider pricing. Figures are measured, not estimated. All releases · Live pricing · Latest in AI
