We’re excited to announce that, in collaboration with Professor Lingming Zhang from UIUC, we will be hosting the LLM4Code & Trae AI IDE Reception during #ICSE2025 in Ottawa, Canada, tentatively scheduled for May 2nd. This dinner will be a great opportunity to network with

Our paper has been accepted to ACL 2025 '25 Main Conference! In this paper, we present subtask-oriented reinforced fine-tuning (SoRFT) that significantly enhances LLMs' issue-resolving performance, which is powering the in-house model of TRAE ! 🔗 arxiv.org/abs/2502.20127

Trae Agent 2.0 just achieved #1 on SWE-bench Verified with Claude 3.7, reaching a 71.0% accuracy. swebench.com We'll continue pushing the boundaries of coding with Claude 4.0 and more. 🧵Here's how we achieved this success on the industry's toughest benchmark:

Cool Details from TRAE on how they build the best Agent on SWE-bench Verified with a 70.6% accuracy using Claude 3.7 using a sophisticated multi-stage agent system and 4 tools. Employs a multi-agent process: 1. Coder agents to create multiple patch candidates. 2. Tester

🤔 Have we really made great progress on software engineering tasks? 🚀 Introducing SWE-bench-Live, a live-updatable benchmark for real-world bug fixing. 😺 Even the best combo, OpenHands + Claude 3.7 Sonnet, sees a major performance drop! 👉 swe-bench-live.github.io 🧵 1/4

Trae just hit #1 on SWE-Bench with Claude 4. We've seen a lot of improvements in real-world coding tasks and complex bug fixes. Claude 4 on Trae is now officially the strongest performer. Huge respect to the Anthropic team for building such an incredible foundation. 🙌

After weeks of hard work with my colleagues Pengfei Gao and Zhao Tian, we’ve reached #1 on SWE-bench Verified and open-sourced our solution! 🚀 We’re excited to collaborate with the OSS community to make it even better and easier to use. All kinds of contributions are welcome!

🎉 Excited to share our latest work from Trae Research Team: Trae Agent: An LLM-based Agent for Software Engineering with Test-time Scaling! We introduce the first agent-based ensemble reasoning framework for repository-level issue resolution, addressing challenges of large

🚀 Excited to introduce ToolTrain, a tool-integrated training framework that supercharges LLMs for deep repo search and issue localisation. Using a combo of supervised fine-tuning + RL, ToolTrain outperform larger proprietary models on function-level localisation. #LLM

昨天去 TRAE Hackathon，发现 TRAE 不止在做 IDE， 还整了个开源 AI Agent 项目——Trae Agent🤖 简单说，就是： 🛠️ 会写代码、改文件、跑命令、做推理 🚀 一次接好多个 LLM（OpenAI / Claude / Gemini / 本地 Ollama…） 📜 全流程能记录，可复现实验，想折腾架构的人超爽 🧩

Had an amazing day at Buildathon ⚡—great talk with Andrew Ng on the future of AI-assisted coding, and a fun panel discussion with Michele Catasta from Replit ⠕, Paxton Maeder-York from Anthropic and Eli Chen from AI Fund on best practices for building with AI. Exciting to see 100+ devs

Best AI hackathon vibe with AI Fund🔥 - TRAE SOLO blew devs away 🤯 - Pitch and Andrew Ng how Trae speeds up everything from idea → deploy (p1) - Our research scientist Chao Peng dropped share our insight of SWE-Bench #1 win with Replit ⠕ CEO + Paxton Maeder-York

🎉 Thrilled to share our new papers from Trae Vassallo Research at NeurIPS Conference, ICSE, and ASE 2024 — exploring how AI coding agents evolve from capable assistants to autonomous developers, spanning execution, evaluation, and model training. [NeurIPS Spotlight] Repo2Run — first

⏰ 3 days left to submit to #LLM4Code2026 (co-located with ICSE 2026)! Pls submit by Oct 31: llm4code.github.io We hope to create a platform for researchers/practitioners to discuss the latest in LLMs/Agents for Code across diverse fields: SE, PL, ML, and beyond.

11.12.2025 TRAE SOLO goes GA! Global release. Completely FREE to use for a limited time！

Tired of comparing LLMs across proprietary, apples-to-oranges agent scaffolds for SWE tasks? 📢📢 Introducing a unified leaderboard: all models are evaluated using live-SWE-agent, the first live software agent that self-evolves on the fly. 🔥 Opus 4.5 + live-SWE-agent hits

Excited to share our latest work on dynamic turn control for LLM-based Agents 🚀 This is the first systematic study that significantly reduces Coding Agent cost with minimal impact on performance, and even improves solve rates in certain scenarios. Our method is intentionally