Open-Source AI

Today's Stories

1. 1

   Alibaba abandons open-source strategy with new Qwen3.7-Plus AI model

   Alibaba released Qwen3.7-Plus, an AI model that can process text, images, and video content at 60% lower cost than its predecessor. However, the company made this version proprietary (requiring paid access) instead of open-source (free for anyone to use), marking a major shift from its previous strategy of releasing powerful models for free.

   Companies like Airbnb that relied on free Qwen models will need to either pay for the new version or find alternative AI solutions for their products.

2. 2

   Microsoft releases open-source testing framework for AI applications

   Microsoft launched Adaptive Spec-driven Scoring for Evaluation and Regression Testing, an open-source tool that lets developers create AI behavior tests using simple text descriptions. The framework helps ensure AI systems work correctly before being deployed in real applications.

   This could lead to more reliable AI features in the apps and services people use daily, as developers can more easily catch problems before release.

3. 3

   New interactive tool helps match AI models to graphics cards

   AgentSwarms published an interactive guide that gamifies the process of figuring out which graphics cards (GPUs) can run which AI models. Users can select model sizes and settings to instantly see what hardware they need, replacing static technical guides.

   Small businesses and hobbyists can now more easily determine if they can run AI models on their existing computers without expensive trial and error.

4. 4

   Users achieve impressive speeds running large AI models on older gaming cards

   Reddit users report successfully running Qwen 3.6 (a 35-billion parameter AI model) on a five-year-old NVIDIA RTX 3090 graphics card, achieving 160 tokens per second output speed. The model was compressed to fit in the card's 24GB memory and can process both text and images.

   People can run sophisticated AI capabilities locally on older gaming hardware they may already own, without paying for cloud AI services.

5. 5

   Google's LiteRT engine shows 2.4x speed improvement over standard AI formats

   Tests comparing Google's LiteRT format with standard GGUF format for running Gemma 4 AI models showed significant speed improvements in text generation, though image processing speeds remained similar. The tests used identical prompts and measured output tokens per second.

   AI applications could become noticeably faster and more responsive as developers adopt more efficient model formats like LiteRT.