2 min read

OpenAI’s Sora 2 Redefines Video AI as AMD Tests 20+ Local Coding Models

New Models and Releases

Sora 2 by OpenAI: OpenAI released Sora 2, a major upgrade to its video generation model, featuring enhanced realism and creative capabilities, including anime-style content and surreal scenes.

VibeVoice 1.8.0: Fabio Sarracino released the first 8-bit quantized VibeVoice model, reducing generation time while maintaining quality. The update includes dynamic quantization and a new manual model management system.


Model Evaluations and Benchmarks

Anthropic’s Sonnet 4.5 and Zhipu-AI’s GLM-4.6: Sonnet 4.5 leads EQ-Bench creative writing evaluations, while GLM-4.6 shows incremental improvements. Zhipu-AI also open-sourced their RL post-training framework, Slime.

AMD’s Local Model Testing: AMD tested 20+ local AI models for coding, finding only Qwen3-Coder and GLM-4.5-Air performed consistently well, emphasizing the role of quantization and hardware compatibility.


Open-Source and Community Developments

Minecraft LLM Experiment: A developer created a small language model (5M parameters) within Minecraft, trained on the TinyChat dataset, without using command blocks or datapacks.

Open-Source Funding Crisis: A developer spent $3K on community-voted open-source models (e.g., Art-1-20B, AGI-0) but faces financial constraints, seeking sustainable funding solutions like sponsorships or paid access.


Partnerships and Academic Collaborations

King’s College London × Mistral AI: The university partnered with Mistral AI to give students exclusive access to Le Chat, enabling customization and deployment of generative AI models for educational applications.


AI in Software Development

Prototype-First Design with AI Agents: An article explores using AI agents for rapid prototyping in software development, improving efficiency and creativity in the design process.