Alibaba’s New AI Coder Surpasses DeepSeek

Alibaba has launched what it calls the “most advanced AI tool for programming”—an open-source AI model named Qwen3-Coder.

>>> Qwen3-Coder is here! ✅

We’re releasing Qwen3-Coder-480B-A35B-Instruct, our most powerful open agentic code model to date. This 480B-parameter Mixture-of-Experts model (35B active) natively supports 256K context and scales to 1M context with extrapolation. It achieves… pic.twitter.com/Z8HfyrVScE

— Qwen (@Alibaba_Qwen) July 22, 2025

The neural network is designed for tasks related to programming. It can write code and manage complex workflows, showing progress in other areas:

• following instructions;
• logical reasoning;
• text comprehension;
• mathematics;
• natural sciences;
• tool usage.

The company positions the product as “particularly strong” in agentic AI tasks—where artificial intelligence can operate independently without human intervention.

pic.twitter.com/M8krjTM4tq

— Qwen (@Alibaba_Qwen) July 22, 2025

According to the data presented, Qwen3-Coder has outperformed Chinese competitors, including models from DeepSeek and K2 by Moonshot AI, across several metrics. In the American Mathematics Olympiad, it scored 70.3 points compared to DeepSeek-V3-0324’s 46.6.

In the MultiPL-E programming test, the language model scored 87.9 points, while DeepSeek scored 82.2.

In some areas, the neural network is comparable to leading American counterparts such as Claude from Anthropic and GPT-4 from OpenAI.

The Qwen3-235B-A22B-Instruct-2507-FP8 neural network is an update to the previous version Qwen3-235B-A22B-FP8, but operates in a non-reasoning mode—producing results without step-by-step logical deductions. This allows the maximum length of processed content to increase eightfold to 256,000 tokens, enabling work with much larger texts within a single dialogue.

Back in May, Alibaba announced a breakthrough in reducing the cost of training AI models focused on search.