Qwen3-Coder: Alibaba’s 480‑Billion‑Parameter Open‑Source Code Model Takes on Claude 4

Introduction

Alibaba’s Qwen team recently launched Qwen3-Coder, the most advanced code‑focused large language model (LLM) in the open‑source ecosystem. The flagship 480 billion‑parameter version (referred to as 480 B) sets a new benchmark for code generation, browser reasoning, and tool interaction, with performance that approaches Claude 4 Sonnet.

What Is Qwen3-Coder?

Qwen3-Coder is the code‑specialized branch of the Qwen3 series. It is built for long‑context intelligent programming, using YaRN technology to support up to 1 million tokens (actual activation parameters: 35 billion). The model inherits the strengths of the Qwen3 base model, excelling not only in code generation but also in mathematics, logical reasoning, and general language tasks. It supports 358 programming languages, including Python, JavaScript, TypeScript, C++, Java, Go, Rust, PHP, SQL, HTML, CSS, and Shell.

Architecture and Training Details

Model type: causal language model

Training stages: pre‑training and post‑training

Total parameters: 480 billion, active parameters: 35 billion

Layers: 62

Attention heads (grouped‑query): 96 query heads, 8 key/value heads

Experts: 160 total, 8 active per forward pass

Native context length: 262,144 tokens (extended to 1 M via YaRN)

The pre‑training phase consumed 7.5 trillion tokens, with 70 % of the data sourced from code. The team refined noisy samples using the Qwen2.5‑Coder pipeline, which markedly improved data quality and reduced hallucinations. Post‑training focuses on reinforcement learning (RL) and a specialized Code RL curriculum, coupled with execution‑based verification to boost the model’s ability to generate runnable code. Long‑horizon RL was also applied to improve performance on multi‑step planning benchmarks such as SWE‑Bench.

Performance Evaluation

According to Qwen’s published benchmarks, Qwen3‑Coder‑480B‑A35B‑Instruct currently leads all open‑source intelligent agents. It outperforms other models on coding tasks, browser‑based reasoning, and tool usage, achieving scores of 37.5 on Terminal‑Bench, 55.8 on Mind2Web, and 77.5 on TAU‑Bench Retail. However, it still trails Claude Sonnet‑4 on complex tasks, where Claude demonstrates higher control and accuracy on benchmarks like SWE‑Bench and BFCL.

Accessing the Model

The full model weights are publicly available on GitHub. Cloning the repository downloads multiple Safetensor files, each roughly 9‑10 GB, so sufficient local storage is required.

git clone https://github.com/QwenLM/qwen-code.git
cd qwen-code && npm install && npm install -g

Qwen also provides a command‑line tool called Qwen Code , derived from Gemini Code, which adapts prompts and function‑call protocols to fully exploit Qwen3‑Coder’s capabilities.

export OPENAI_API_KEY="your_api_key_here"
export OPENAI_BASE_URL="https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
export OPENAI_MODEL="qwen3-coder-plus"

API and Pricing

The commercial variant Qwen3‑coder‑plus offers API access with tiered pricing based on token consumption. Key limits include a 1,048,576‑token context window, a maximum input of 1 million tokens, and a maximum output of 65,536 tokens. A free quota of 1 million tokens is available for 180 days after activation. Pricing scales from $1 per million input tokens (0‑32 K) up to $6 per million input tokens (256 K‑1 M), with corresponding output rates.

Example Usage

Using the OpenAI‑compatible API, the model can generate a Python function to find prime numbers:

import os
from openai import OpenAI
client = OpenAI(api_key=os.getenv("DASHSCOPE_API_KEY"), base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1")
completion = client.chat.completions.create(
    model="qwen3-coder-plus",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Please write a Python function find_prime_numbers that takes an integer n as a parameter and returns a list containing all prime numbers less than n. Do not output non‑code content."}
    ]
)
print(completion.choices[0].message.content)

The response correctly returns a complete Python implementation. Additional demos include generating a full‑stack web app via a HuggingFace space and creating SVG‑based Pokémon illustrations in under 30 seconds.

Conclusion

Qwen3‑Coder provides a powerful, fully open‑source alternative to proprietary code assistants. Its extensive benchmarks, long‑context capability, and open accessibility make it a compelling option for developers seeking a high‑performance coding assistant without vendor lock‑in.