GLM-Image & Abliterated Llama 3.3 Debut as LoongFlow AI Agents Self-Evolve
New AI Models & Releases
GLM-Image Model Announced by Z.ai: A new multimodal model, GLM-Image, has been spotted in Hugging Face’s Transformers repository, extending the GLM series to handle image-related tasks. The model’s architecture is defined via a configuration class, but full details remain limited.
Abliterated Llama 3.3 8B Released: A modified version of Llama 3.3 8B has been optimized for compliance while minimizing intelligence loss, achieving a KL divergence of <0.05. The model is available in BF16 weights on Hugging Face.
AI Research & Frameworks
LoongFlow: Self-Evolving AI Agents Framework: Researchers introduced LoongFlow, a framework using evolutionary algorithms to optimize AI agents by treating prompts and logic as "DNA." The paper and code are publicly available, with the team seeking community feedback.
Adaptive-P Sampler for Creative Text Generation: A new Adaptive-P sampling method for llama.cpp aims to break repetitive text generation by dynamically adjusting token selection probabilities. It’s designed for creative tasks like fiction, roleplay, and brainstorming.
AI Tools & Workflows
EmergentFlow: Browser-Based Visual AI Workflow Tool: A new no-code tool, EmergentFlow, allows users to build AI workflows visually in-browser, supporting local models (Ollama, LM Studio, llama.cpp) and cloud APIs (OpenAI, Anthropic, Groq, etc.). It includes agents, web search, TTS, and more.
Community & Milestones
vLLM Reaches 2,000 Contributors: The open-source vLLM project hit a milestone of 2,000 contributors, reflecting strong community engagement in optimizing LLM inference and serving.
Industry Insights
OpenAI’s Greg Brockman & Django’s Simon Willison on AI-Assisted Coding: In a discussion, the two highlighted how models like Opus 4.5 and GPT-5.2 are transforming software engineering by handling larger codebases and complex tasks, reducing manual coding needs.