Gamechanger: Qwen3 released!

The AI race reaches new heights! With the release of Qwen3, a groundbreaking concept is establishing itself in the world of large language models: language models that can switch between quick responses and deep thinking. The flagship Qwen3-235B-A22B successfully competes with industry giants such as DeepSeek-R1, o1, and Gemini-2.5-Pro – and does so with significantly more efficient use of resources.

The innovation lies in the hybrid approach: in “Thinking Mode,” the model works step by step on complex problems, while in “Non-Thinking Mode,” quick answers are provided for simpler questions. This flexibility enables demand-driven control of “thinking power” and optimizes the balance between efficiency and quality.

With support for 119 languages and improved capabilities in the areas of code and agent interaction, Qwen3 is positioning itself as a versatile tool for developers and researchers. Particularly noteworthy: The family includes both MoE (mixture-of-experts) models and dense models of various sizes – all available under the Apache 2.0 license.

Will this hybrid approach enable us to reach the next milestone on the road to AGI? The combination of scaled pre-training and an innovative thinking concept could be just the right step.

Why it matters: Qwen3's hybrid thinking architecture marks a significant advance for practical AI applications, as it is the first to truly combine on-demand thinking with scalable performance. This balance between computational efficiency and depth of thought could pave the way for more responsible, accessible AI systems that are capable of both rapid response and deep analysis.

This is the easiest way for a busy person wanting to learn AI in as little time as possible:

Sign up to start learning.

“To develop the hybrid model capable of both step-by-step reasoning and rapid responses, we implemented a four-stage training pipeline. This pipeline includes: (1) long chain-of-thought (CoT) cold start, (2) reasoning-based reinforcement learning (RL), (3) thinking mode fusion, and (4) general RL.” (Qwen 3)

Researchers have conducted a comprehensive analysis of “sparse attention” in LLMs – a technique for processing long texts more efficiently. What is new is the systematic investigation of the trade-offs between speed and accuracy across many models and tasks. The key finding: for very long sequences, large, highly “sparse” models outperform smaller, dense models with the same computing budget. However, sparseness is not a universal solution and carries performance risks for specific tasks. It is relevant for processing huge amounts of data, but requires careful, application-specific consideration.

Automatically summarize long videos – without expensive data? ViSMaP makes it possible! This new AI system cleverly uses multiple language models (LLMs) via “meta-prompting.” It generates amazingly good, automatic summaries for hours of video from descriptions of short clips – without any human annotation. The performance is comparable to that of elaborately trained methods. This is relevant for analyzing huge video archives and complex content, from social media to film, and circumvents previous data bottlenecks

AI should generate images based on text and a specific template image (e.g., “draw this cat on a tree”). However, evaluating such models is difficult. REFVNLI is a new, cost-efficient AI metric that automatically evaluates both: Was the text description implemented? And: Is the correct subject (e.g., exactly this cat) recognizable? What is new is the combined, reliable evaluation in a single step. This enables advances in personalized image generation and consistent AI creations, which are important for media and design.

Hugging Face has launched the SO-101, a $100 programmable, 3D-printable robotic arm that builds on the success of last year’s SO-100. Developed with The Robot Studio and other partners, the SO-101 is easier to assemble and features upgraded motors and a camera for reinforcement learning-based task training. It can perform simple chores like picking and placing objects, making it ideal for AI hobbyists and developers. The launch comes as Hugging Face expands deeper into robotics, including the acquisition of Pollen Robotics and plans to distribute its humanoid robot, Reachy 2.