• Ainsider
  • Posts
  • The end of April in AI - Ainsider AI Newsletter

The end of April in AI - Ainsider AI Newsletter

The most important updates and upgrades: Qwen | Perplexity | Grok | Leonardo AI | Lovable

Author

Piotr Macai
April 30, 2025

Table of Contents

Alibaba’s Qwen 3: A Game-Changing Open-Source Multimodal LLM Outperforming o3-mini and DeepSeek-R1

Alibaba Cloud has unveiled Qwen 3, a groundbreaking open-source multimodal LLM that surpasses OpenAI’s o3-mini and DeepSeek-R1 in key benchmarks. Released on April 28, 2025, with an Apache 2.0 license, Qwen 3 offers eight models ranging from 0.6B to 235B parameters, catering to diverse computational needs. This versatile AI can generate text, images, apps, and videos—all for free via Qwen Chat—making it a powerful tool for developers and creators worldwide.

  • Flagship Performance: The Qwen3-235B-A22B model excels with scores of 94.0 on ArenaHard and 81.7 on MMLU, outperforming competitors in coding, math, and reasoning tasks [Web ID: 0].

  • Efficient MoE Models: Qwen3-30B-A3B, a Mixture of Experts model, outstrips QwQ-32B while using 10 times fewer activated parameters, optimizing resource-heavy workflows.

  • Compact Power: Even the smallest Qwen3-4B rivals the performance of Qwen2.5-72B-Instruct, ideal for users with limited hardware [Web ID: 6].

  • Global Reach: Supporting 119 languages and dialects, from Indo-European to Sino-Tibetan, Qwen 3 enables seamless international applications.

  • Multimodal Innovation: Beyond text, Qwen 3’s capabilities in image, app, and video generation within Qwen Chat open new creative avenues for users.

  • Scalable Reasoning: Performance scales with computational budgets, allowing users to balance cost and quality, especially in reasoning-heavy tasks.

  • Legacy of Progress: Building on Qwen 2’s multimodal advancements like Qwen2-VL, Qwen 3 continues Alibaba’s open-source AI leadership [Web ID: 2].

  • Seamless Integration: Frameworks like SGLang, vLLM, and tools such as Ollama and MLX ensure easy deployment across research and production environments.

  • Competitive Landscape: Qwen 3 joins 2025’s open-source AI wave, rivaling models like Llama 3.1 and DeepSeek-V3 in scale and multimodal prowess [Web ID: 0].

Qwen 3 sets a new standard for open-source AI, empowering global innovation with its unmatched versatility and performance. Explore it today on Qwen Chat!

Huge Perplexity Update

Perplexity announced a package of significant updates that enhance the user experience, especially on the iOS platform.

The latest version introduces a Voice Assistant on iOS, enabling users to perform various tasks such as booking tables, sending emails, managing calendars, and playing multimedia, all via voice commands.

Another new feature is GPT-powered image generation, allowing users to create visuals based on their queries.

Additionally, Perplexity has introduced new AI models: Grok 3 in beta and o4-mini, available in the model selector, giving users greater flexibility when working with the tool.

These updates confirm that Perplexity is constantly striving to expand its capabilities, becoming an increasingly versatile AI assistant.

New Grok Capabilities & Mobile App

xAI has presented new features for Grok, including a mobile application that significantly increases the accessibility of this powerful AI assistant.

Grok, designed with maximum usefulness and truthfulness in mind, now offers the ability to generate images, answer questions about uploaded photos, and provide detailed answers to a variety of queries.

The mobile application allows users to take advantage of these features anywhere, making Grok an ideal tool for both work and everyday tasks.

Additionally, new features include advanced contextual search, which allows for a better understanding of complex questions and the provision of precise answers.

Flux in Leonardo AI: New Standards in Image Generation

Leonardo AI has introduced the Flux model - in Flux Schnell and Flux Dev variants within its application.

The model offers a wide range of configuration options, including adjusting contrast, resolution (up to 1472x832 pixels), and the number of images generated (up to 4 simultaneously).

Additionally, Flux supports features such as enabling and disabling prompt enhancement, allowing for greater control over the creative process.

Lovable released version 2.0, taking Vibe Coding to the next level 👀 Create apps and websites with:

  • vibe coding with your team 👀

  • 10x smarter chat agent,

  • new user interface,

  • security scans,

  • custom domains

Dreamina AI Seedream 3.0: Leader in Photorealistic Images

Dreamina AI has launched Seedream 3.0, which took first place in generating photorealistic images at resolutions up to 2K. This model also offers features such as upscaling, inpainting, image expansion, and video generation, making it a comprehensive tool for visual creators.

Genspark AI Slides: A Revolution in Presentation Creation

Genspark has presented AI Slides, a tool that changes the approach to presentation creation through the use of an agent-based AI model. The system automatically researches selected topics, generates relevant visualizations such as images and charts, and also converts various types of documents into professional slides. This solution is ideal for individuals seeking efficient and modern ways to prepare presentations.

OpenAI GPT-Image-1: Image Generation Available for Developers

OpenAI has made the GPT-Image-1 model available to developers through its API, enabling the integration of image generation features into external applications. This model, which gained popularity thanks to ChatGPT, generated over 700 million images in the first week after its release in March. This is another step in the development of multimodal AI technologies.

Tavus: State-of-the-Art Lip Sync Model

Tavus has presented a new lip sync model that achieves unprecedented quality in speech and video synthesis. This technology ensures perfect synchronization of lip movements with audio, as well as natural facial expressions, opening up new possibilities in video production and animation.

Nari Labs Dia 1.6B: AI with Advanced Emotional Expression

Nari Labs has introduced Dia 1.6B, an AI model distinguished by its wide emotional range and naturalness. The system can express emotions such as laughter or coughing, making it more human-like in interactions. The model is available to try out on the HuggingFace platform, encouraging experimentation with its capabilities.

ChatGPT Deep Research Mini: Lightweight Version from OpenAI

OpenAI has introduced ChatGPT Deep Research Mini, a lighter version of its research tool that retains key features with lower resource requirements. If the limits of the original version are exceeded, queries will be automatically directed to this more resource-efficient alternative, increasing accessibility for a wider range of users.