AI Newsletter

AI Newsletter

Share this post

AI Newsletter
AI Newsletter
🤖 AI Agents Weekly: Lovable Agents, GitHub Spark, Qwen3-Coder, Search Arena, Awesome Context Engineering

🤖 AI Agents Weekly: Lovable Agents, GitHub Spark, Qwen3-Coder, Search Arena, Awesome Context Engineering

Lovable Agents, GitHub Spark, Qwen3-Coder, Search Arena, Awesome Context Engineering

Jul 26, 2025
∙ Paid
18

Share this post

AI Newsletter
AI Newsletter
🤖 AI Agents Weekly: Lovable Agents, GitHub Spark, Qwen3-Coder, Search Arena, Awesome Context Engineering
1
Share

In today’s issue:

  • Qwen3-Coder

  • Microsoft presents GitHub Spark

  • Benefits of structural planning for LLM Agents

  • Anthropic shares research on building and evaluating alignment auditing agents

  • Awesome Context Engineering repo

  • Lovable announced its new coding agents

  • LMArena has launched Search Arena

  • A new evaluation framework tailored for agentic search

  • OpenAI and Gemini teams achieved a gold medal at IMO 2025

  • Top AI papers, research, tool updates, and much more.



Top Stories

Qwen3-Coder

Qwen3-Coder is a new frontier in agentic code generation from the Qwen team, introduced as a 480B-parameter Mixture-of-Experts model with 35B active parameters. It supports a massive 256K-token native context window (up to 1M with extrapolation) and delivers state-of-the-art performance in agentic coding, browser use, and tool use, rivaling Claude Sonnet 4 among open models. Alongside the model, the team released Qwen Code, a command-line interface for agentic coding, and enabled integration with Claude Code.

Key innovations include:

  • Massive-scale pretraining on 7.5T tokens with a 70% code ratio, enhanced by synthetic data cleanup using earlier Qwen2.5-Coder checkpoints. This retains general and mathematical capabilities while improving code understanding.

  • Code RL at scale, embracing execution-driven reinforcement learning to boost real-world coding task performance. The team scaled test case generation and model tuning to improve execution success rates across diverse tasks.

  • Long-horizon RL (Agent RL) support enables multi-turn interactions with tools and environments (like SWE-Bench). A novel infrastructure allows running 20,000 parallel environments using Alibaba Cloud, unlocking high-throughput agent training and eval.

  • Strong ecosystem tooling: Qwen Code (a Gemini Code fork), OpenAI-compatible API access, and optional Claude Code integration provide multiple ways to deploy and interact with Qwen3-Coder, including support for .env setup, router customization, and direct Cline integration.

Blog | GitHub

Keep reading with a 7-day free trial

Subscribe to AI Newsletter to keep reading this post and get 7 days of free access to the full post archives.

Already a paid subscriber? Sign in
© 2025 elvis
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share