Hello AI Enthusiasts! Welcome to the Nineteenth edition of "This Week in AI Engineering"! "This Week in AI Engineering" This week, Meta introduced AssetGen 2.0, marking a big step in AI-driven 3D modeling, Alibaba’s Qwen2.5 goes open-source with serious upgrades, DeepMind’s AlphaEvolve rewrites the rules of algorithm design, and OpenAI rolls out tools that could redefine how devs interact with code. With this, we'll also explore some under-the-radar tools that can supercharge your development workflow. Alibaba Qwen 3 is a Web Developer's Dream Alibaba Qwen 3 is a Web Developer's Dream Alibaba's latest release, Qwen3, introduces a hybrid thinking architecture combining Mixture of Experts (MoE) models with enhanced reasoning capabilities. Pre-trained on ~36 trillion tokens (roughly double Qwen2.5’s data) spanning 119 languages/dialects Qwen3 Qwen3 Model Specifications: Qwen3-235B-A22B: A 235 billion parameter model optimized for coding, mathematics, and general reasoning tasks.
Performance: Achieves competitive results in benchmark evaluations, rivaling models like DeepSeek-R1 and Gemini-2.5-Pro.
Web Development Focus: Excels in frontend development tasks, translating design specifications into responsive and aesthetically pleasing UIs Model Specifications: Qwen3-235B-A22B: A 235 billion parameter model optimized for coding, mathematics, and general reasoning tasks. Model Specifications Performance: Achieves competitive results in benchmark evaluations, rivaling models like DeepSeek-R1 and Gemini-2.5-Pro. Performance Web Development Focus: Excels in frontend development tasks, translating design specifications into responsive and aesthetically pleasing UIs Web Development Focus translating design specifications translating design specifications In benchmarks, Qwen3 “surpasses previous Qwen” on math, coding, and reasoning tests. For example, Qwen3-30B-A3B (30B MoE) outperforms a 32B QwQ model despite having 10× fewer active params, and even a 4B Qwen3 rivals Qwen2.5-72 B. Overall, the dense bases match or exceed larger Qwen2.5 models on STEM/coding . Google AlphaEvolve Literally Thinks Different Google AlphaEvolve Literally Thinks Different Google DeepMind's AlphaEvolve is pushing the boundaries of algorithm design, surpassing human-devised methods in efficiency. Google DeepMind's AlphaEvolve Google DeepMind's AlphaEvolve Matrix Multiplication Breakthrough: Discovered a method to multiply 4×4 matrices using just 47 steps, improving upon the 1969 Strassen algorithm's 49 steps. It also improves the state of the art for 14 matrix multiplication algorithms over DeepMind’s prior specialized AlphaTensor
Applications: Optimized solutions for data center scheduling, chip design, and language model efficiency.
Significance: Demonstrates AI's capability to generate novel and provably correct algorithms, marking a milestone in AI-driven innovation. Matrix Multiplication Breakthrough: Discovered a method to multiply 4×4 matrices using just 47 steps, improving upon the 1969 Strassen algorithm's 49 steps. It also improves the state of the art for 14 matrix multiplication algorithms over DeepMind’s prior specialized AlphaTensor Matrix Multiplication Breakthrough: Applications: Optimized solutions for data center scheduling, chip design, and language model efficiency. Applications: Significance: Demonstrates AI's capability to generate novel and provably correct algorithms, marking a milestone in AI-driven innovation. Significance In testing over 50 open problems in math and CS, it rediscovered known solutions ~75% of the time and improved ~20% (e.g. solving the 11-dimensional “kissing number” problem with 593 spheres vs. the old record of 592) Meta's AssetGen 2.0 Generates Top-Tier 3D Models Meta's AssetGen 2.0 Generates Top-Tier 3D Models Meta has unveiled AssetGen 2.0, a significant leap in AI-driven 3D content creation. This single-stage 3D diffusion model generates high-fidelity meshes directly from text prompts, eliminating the need for intermediate representations. AssetGen 2.0 AssetGen 2.0 TextureGen Integration: AssetGen 2.0 integrates TextureGen, which applies high-resolution, view-consistent textures using physically-based rendering (PBR) materials. This integration ensures that the generated 3D assets are not only geometrically accurate but also visually realistic.
Training Data: Trained on an extensive corpus of 3D assets, ensuring diverse and accurate outputs. While Meta has not publicly disclosed the specific datasets or types of 3D assets used to train AssetGen 2.0. The available information indicates that AssetGen 2.0 was trained on a large corpus of 3D assets to enhance the diversity and accuracy of its outputs TextureGen Integration: AssetGen 2.0 integrates TextureGen, which applies high-resolution, view-consistent textures using physically-based rendering (PBR) materials. This integration ensures that the generated 3D assets are not only geometrically accurate but also visually realistic. TextureGen Integration Training Data: Trained on an extensive corpus of 3D assets, ensuring diverse and accurate outputs. While Meta has not publicly disclosed the specific datasets or types of 3D assets used to train AssetGen 2.0. The available information indicates that AssetGen 2.0 was trained on a large corpus of 3D assets to enhance the diversity and accuracy of its outputs Training Data Application: Already used internally by Meta to create VR world content and Horizon/Avatar assets. It will soon roll out to Meta Horizon creators (via Horizon Desktop Editor) later in 2025, Meta envisions using AssetGen 2.0 as a building block for auto‑generating entire 3D scenes Application Efficient Language Modeling with IBM's Bamba-9B v2 Efficient Language Modeling with IBM's Bamba-9B v2 Training Enhancements: Trained on an additional 1 trillion tokens, significantly improving performance over its predecessor.
Benchmark Performance: On standard NLP benchmarks (L1/L2 leaderboards), Bamba-9B-v2 outperforms Meta’s Llama 3.1 8B (which was trained on ~7× more data) Bamba-9B also supports very long context: trained on 4K sequences but can handle up to 32K tokens, with potential for 100K+ as vLLM adds better SSM support.
Deployment: Bamba-9B-v2 is fully open-source (Apache 2.0) on Hugging Face, offering flexible deployment options, including quantization for efficient inference. Training Enhancements: Trained on an additional 1 trillion tokens, significantly improving performance over its predecessor. Training Enhancements: Benchmark Performance: On standard NLP benchmarks (L1/L2 leaderboards), Bamba-9B-v2 outperforms Meta’s Llama 3.1 8B (which was trained on ~7× more data) Bamba-9B also supports very long context: trained on 4K sequences but can handle up to 32K tokens, with potential for 100K+ as vLLM adds better SSM support. Benchmark Performance: Deployment: Bamba-9B-v2 is fully open-source (Apache 2.0) on Hugging Face, offering flexible deployment options, including quantization for efficient inference. Deployment: OpenAI's ChatGPT Integrates GitHub Connector OpenAI's ChatGPT Integrates GitHub Connector OpenAI has introduced a GitHub connector for ChatGPT's Deep Research tool, which allows ChatGPT (GPT-4-based) to securely link to GitHub repositories. Users can ask questions that require reading code, docs, or issues from a GitHub repo, and ChatGPT will retrieve and analyze the relevant content. GitHub connector GitHub connector Features: Features: Natural Language Queries: Users can ask questions about their codebases, and ChatGPT will provide context-aware answers.
Code Summarization: Generates summaries of functions and modules, aiding in understanding complex code structures.
Dependency Mapping: Identifies and visualizes dependencies within the codebase.
Availability: Currently rolling out to ChatGPT Plus, Pro, and Team users, with Enterprise and Education support forthcoming. Natural Language Queries: Users can ask questions about their codebases, and ChatGPT will provide context-aware answers. Natural Language Queries: Code Summarization: Generates summaries of functions and modules, aiding in understanding complex code structures. Code Summarization: Dependency Mapping: Identifies and visualizes dependencies within the codebase. Dependency Mapping: Availability: Currently rolling out to ChatGPT Plus, Pro, and Team users, with Enterprise and Education support forthcoming. Availability ChatGPT Plus Pro Project Kiro: Amazon's New Coding Assistant Project Kiro: Amazon's New Coding Assistant Amazon Web Services (AWS) is developing Project Kiro, an AI-powered coding assistant designed to streamline software development. It is a web/desktop application that orchestrates multiple AI agents (Amazon’s and third-party) along with domain knowledge and extensions to automate software development tasks. Project Kiro Project Kiro Features: Features: Multi-Modal Interface: Accepts inputs in various forms, including text, diagrams, and structured data.
Real-Time Code Generation: Utilizes AI agents to generate code snippets based on user prompts and context.
Integration: Designed to work seamlessly with existing AWS tools and services.
Deployment: According to reports, AWS aimed to beta-launch Kiro around late June 2025, to be available as both a web and desktop application, catering to diverse development workflows. Multi-Modal Interface: Accepts inputs in various forms, including text, diagrams, and structured data. Multi-Modal Interface Real-Time Code Generation: Utilizes AI agents to generate code snippets based on user prompts and context. Real-Time Code Generation: Integration: Designed to work seamlessly with existing AWS tools and services. Integration: Deployment: According to reports, AWS aimed to beta-launch Kiro around late June 2025, to be available as both a web and desktop application, catering to diverse development workflows. Deployment No model sizes or benchmarks are public, as this is an emerging internal system. Tools & Releases YOU Should Know About Tools & Releases YOU Should Know About WebThinker: Autonomous Web Research Agent WebThinker: Autonomous Web Research Agent WebThinker WebThinker empowers large reasoning models to autonomously browse the web, gather real-time information, and generate detailed research reports. Key Features: Key Features: Deep Web Explorer: Enables dynamic search and navigation of web pages.
Autonomous Think-Search-and-Draft: Allows seamless integration of reasoning, information gathering, and report writing.
RL-based Training: Employs reinforcement learning via Direct Preference Optimization for enhanced research capabilities. Deep Web Explorer: Enables dynamic search and navigation of web pages. Deep Web Explorer Autonomous Think-Search-and-Draft: Allows seamless integration of reasoning, information gathering, and report writing. Autonomous Think-Search-and-Draft: RL-based Training: Employs reinforcement learning via Direct Preference Optimization for enhanced research capabilities. RL-based Training DeerFlow: Community-Driven Deep Research Framework DeerFlow: Community-Driven Deep Research Framework DeerFlow Developed by ByteDance, DeerFlow is an open-source framework that combines language models with specialized tools for comprehensive research tasks. Architecture: Built on LangChain and LangGraph, offering a modular and extensible platform.
Capabilities: Supports tasks like web search, crawling, and Python code execution.
Community Focus: Aims to give back to the open-source community by integrating and enhancing existing tools. Architecture: Built on LangChain and LangGraph, offering a modular and extensible platform. Architecture: Capabilities: Supports tasks like web search, crawling, and Python code execution. Capabilities: Community Focus: Aims to give back to the open-source community by integrating and enhancing existing tools. Community Focus: SunaAI: Open-Source Generalist AI Agent SunaAI: Open-Source Generalist AI Agent SunaAI: Kortix AI's SunaAI is a fully open-source AI assistant designed to perform real-world tasks with human-like autonomy. Functionalities: Interacts with virtual systems, writes files, executes code, and browses the internet.
Deployment: Available under the Apache 2.0 license, supporting both cloud and self-hosted environments.
Use Cases: Ideal for research, data analysis, and automating everyday tasks. Functionalities: Interacts with virtual systems, writes files, executes code, and browses the internet. Functionalities: Deployment: Available under the Apache 2.0 license, supporting both cloud and self-hosted environments. Deployment: Use Cases: Ideal for research, data analysis, and automating everyday tasks. Use Cases: DocuWriter.ai: Automated Code & API Documentation DocuWriter.ai: Automated Code & API Documentation DocuWriter.ai DocuWriter.ai is an AI-powered web application that generates automated code and API documentation from your source code files. Features: Features: Code Comments & DocBlock Generator: Automatically adds descriptive comments to your code.
UML Diagram Generator: Visualizes code structure for better understanding.
AI-Powered Code Tests Suite Generation: Creates test suites to ensure code reliability.
Intelligent Code Refactoring: Suggests improvements for cleaner and more efficient code. Code Comments & DocBlock Generator: Automatically adds descriptive comments to your code. Code Comments & DocBlock Generator: UML Diagram Generator: Visualizes code structure for better understanding. UML Diagram Generator: AI-Powered Code Tests Suite Generation: Creates test suites to ensure code reliability. AI-Powered Code Tests Suite Generation: Intelligent Code Refactoring: Suggests improvements for cleaner and more efficient code. Intelligent Code Refactoring: And that wraps up this issue of "This Week in AI Engineering." This Week in AI Engineering. Thank you for tuning in! Be sure to share this newsletter with your fellow AI enthusiasts and follow for more weekly updates. Until next time, happy building!

Amazon

Alibaba Qwen 3 Could Be a Web Developer's Dream

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Replit AI Coding Assistant Deletes Company Database

I Stopped Using ChatGPT to Write Code. Here Is What Happened to My Brain.

Stop Asking AI for Answers: Why the Future of Research is "Problem Generation"

Intelligent Document Processing: A Simple Chunking Strategy for AI Content Generation

The Intoxication—and Limits—of AI-Assisted Development

Stop Blindly Building AI Swarms: The New "Scaling Laws" for Agents Are Here

Replit AI Coding Assistant Deletes Company Database

I Stopped Using ChatGPT to Write Code. Here Is What Happened to My Brain.

Stop Asking AI for Answers: Why the Future of Research is "Problem Generation"

Intelligent Document Processing: A Simple Chunking Strategy for AI Content Generation

The Intoxication—and Limits—of AI-Assisted Development

Stop Blindly Building AI Swarms: The New "Scaling Laws" for Agents Are Here

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps