AI
Alibaba Sets New Standards in the Open-Source AI Market – Qwen3-Coder Surpasses GPT‑4o in Coding Benchmarks
With Qwen3-Coder, Alibaba is offering for the first time an agent-based open-source model that significantly surpasses GPT-4o and DeepSeek.
With the launch of Qwen-3 Coder on July 23, 2025, Alibaba pushes to the forefront of the global coding AI market. The model, freely available under the Apache 2.0 license, combines 480 billion parameters (35 billion of which are active per query) with a context depth of up to one million tokens – more than twice as many as GPT-4o.
Trained on 7.5 trillion tokens, approximately 70% of which are code, Qwen3-Coder specializes in agent functions: In addition to reinforcement learning for verifiable code execution, Alibaba uses a system for long-term planning and adaptation over multiple interaction steps. The training architecture is based on a mixture-of-experts approach with 160 experts, of which eight are active per request.
In benchmarks, the model delivers above-average results. On SWE-bench Verified, Qwen3-Coder scored 67.0% (Standard) and 69.6% (500-Turn), compared to 54.6% for GPT-4.1 and 49.0% for Gemini 2.5 Pro. In the MultiPL-E coding test, the model scored 87.9, just behind Claude Opus 4, but clearly ahead of DeepSeek (82.2) and GPT-4o (82.7). In the AIME math exam, the related sister model Qwen3-235B-A22B even significantly outperformed the competition with 70.3 points.
The integration into existing development environments is seamless. Alibaba provides its own CLI tool with structured prompting, Node.js support, and support for systems like Claude Code, Ollama, LMStudio, or llama.cpp. The model can be operated locally, via OpenAI-compatible APIs, or through Alibaba Cloud. The pricing structure is transparent: 1 to 6 USD per million tokens (prompt), depending on context depth – significantly cheaper than closed-source alternatives.
First feedback from the developer community is clear. Sebastian Raschka calls it the "best coding model so far." Wolfram Ravenwolf describes it as "currently unrivaled." Jack Dorsey sees a "new class" of intelligent software development in combination with his agent framework Goose.
The model is already embedded commercially: A slimmed-down 3B version of Qwen3-Coder is set to power HP's smart office assistant Xiaowei Hui in China starting this fall. Thus, the model becomes a central infrastructure component for AI-supported documentation and meeting summaries in the corporate context.
With this step, Alibaba underscores its strategic course towards an open, agent-based, scalable AI ecosystem for businesses — with the clear goal of asserting long-term technological sovereignty over closed systems from the USA.
