By Ann Cao
Alibaba Group Holding鈥檚 open-source Qwen artificial intelligence (AI) model has enabled the agentic framework DeepSWE to outperform rival systems in this growing segment, according to the software platform鈥檚 developers.
Jointly developed by open-source initiative Agentica and San Francisco-based start-up Together AI, DeepSWE was trained on the Qwen3-32B large language model (LLM) 鈥 part of Alibaba Cloud鈥檚 third-generation family of AI models. It topped the leaderboard of the latest SWEBench-Verified test, scoring 59 per cent accuracy against other so-called open-weight models like DeepSeek鈥檚 V3-0324, the developers said in a blog post on Wednesday.
Agentic frameworks are software platforms that provide the structure, tools and functionalities to build, deploy and manage AI agents. They enable AI agents to collaborate, make decisions and automate complex tasks.
AI agents, such as Chinese start-up Butterfly Effect鈥檚 Manus, are software programs that are capable of autonomously performing tasks on behalf of a user or another system. Essentially, these agents create a plan of specific tasks and subtasks to complete a goal using available resources.
DeepSWE marks the latest example of Hangzhou-based Alibaba鈥檚 growing leadership position in the global open-source community. Alibaba owns the South China Morning Post.
The open-source approach gives public access to a program鈥檚 source code, allowing third-party software developers to modify or share its design, fix broken links or scale up its capabilities.
DeepSWE was developed by post-training the Qwen3-32B model using rLLM, Agentica鈥檚 modular reinforcement learning (RL) system.
鈥淲e鈥檝e open-sourced everything 鈥 our data set, code, training and eval logs 鈥 for everyone to progress on scaling and improving agents with RL,鈥 the developers鈥 blog post said.
Trained for six days on a computing facility powered by Nvidia鈥檚 H100 graphics processing units, DeepSWE was specifically trained to solve complex software engineering tasks such as implementing new code features, debugging and solving issues on the online developer platform GitHub.
The development of DeepSWE comes more than two months after Alibaba Cloud made Qwen3 available on more developer platforms online, as the company pushed for wider international adoption of its open-source systems. Alibaba Cloud started open-sourcing Qwen models in August 2023.
Released in April, the Qwen3 AI models can be deployed via LLM platforms Ollama, LM Studio, SGLang and vLLM, according to the Qwen team鈥檚 post on their X account.
Benchmark tests cited by Alibaba in April said models such as Qwen3-235B and Qwen3-4B either matched or exceeded the performance of advanced models from both overseas and domestic competitors 鈥 including ChatGPT creator OpenAI鈥檚 o1, Google鈥檚 Gemini and DeepSeek鈥檚 R1 鈥 in areas like instruction following, coding assistance, text generation, maths skills and complex problem solving.
Qwen was now 鈥渢he world鈥檚 largest open-source model family鈥, Alibaba chairman Joe Tsai and CEO Eddie Wu Yongming said in a letter to shareholders last month. They added that the company had open-sourced more than 200 Qwen models as of April, generating more than 300 million global downloads and over 100,000 derivative models.
Alibaba Cloud on Thursday announced that it would invest more than US$60 million before the end of its current financial year in March to accelerate AI innovation via its partner ecosystem.
That followed Alibaba CEO Wu鈥檚 commitment in February to 鈥渁ggressively invest鈥 in AI and cloud computing infrastructure, with an outlay of at least 380 billion yuan (US$53 billion) over the next three years 鈥 the largest-ever computing project financed by a single private business in China.