Alibaba Unveils Qwen-3 Coder, Its “Most Advanced Coding AI”

Open-source launch

Released on July 23, 2025 under the Apache 2.0 license, Qwen3-Coder is available on HuggingFace, GitHub, Qwen Chat, Alibaba Cloud API, and other channels. Alibaba describes it as its most powerful agentic open-source coding model to date (VentureBeat, SCMP, Mint).

Model architecture

• Qwen3-Coder is a Mixture-of-Experts (MoE) model with 480B total parameters, 35B active per query, and 8 experts selected from 160.
• Features include 62 transformer layers, 96 attention heads, 8 key-value heads, and native support for 256K token context, extended to 1M tokens via YaRN positional extrapolation (VentureBeat, Mint).

Training details

The model is pretrained on 7.5 trillion tokens, ~70% of which are code, and further refined with:

• Code RL (Reinforcement Learning) for verifiable execution tasks

• Long-Horizon Agent RL for planning and adapting over multi-step interactions

  Training involved a 20,000-environment simulation system on Alibaba Cloud (VentureBeat).

Performance and benchmarks

• On SWE-bench Verified, it scored 67.0% (standard) and 69.6% (500-turn).
  • In comparison: GPT-4.1 scored 54.6%, Gemini 2.5 Pro 49.0%, Claude Sonnet-4 70.4% (Medium, SCMP).
• The related Qwen3-235B-A22B-2507-FP8 scored 70.3 on the 2025 American Invitational Mathematics Examination, beating DeepSeek‑V3 (46.6) and GPT‑4o (26.7).
• On the MultiPL-E coding benchmark, Qwen scored 87.9, ahead of DeepSeek (82.2) and GPT‑4o (82.7), trailing only Claude Opus 4 (88.5) (SCMP, Mint).

Tooling & integration

• Qwen Code CLI (open-sourced alongside the model) is forked from Gemini Code. It includes structured prompting, custom function call protocols, and Node.js support.
• Integrates with: Claude Code (DashScope), Cline, Ollama, LMStudio, MLX‑LM, llama.cpp, and KTransformers.
• Works via OpenAI-compatible APIs and can be deployed locally or via Alibaba Cloud (VentureBeat, Mint).

Enterprise adoption & cost

• As a fully open-source model, Qwen3-Coder can be self-hosted with no license fees, enabling vendor-neutral deployments.
• Alternatively, use via Alibaba Cloud costs:

  • $1/$5 per million tokens (up to 32K)

  • $1.8/$9 (128K), $3/$15 (256K), $6/$60 (1M)

    (VentureBeat, Medium).

Early feedback

• Sebastian Raschka: “Best coding model yet.”
• Wolfram Ravenwolf: “This is surely the best one currently.”
• Kevin Nelson: “Qwen3 Coder is on another level.”
• Jack Dorsey: “Goose + qwen3-coder = wow,” referring to integration with his open-source agent framework Goose (VentureBeat).

Future developments

• Smaller Qwen3-Coder variants are in development to reduce inference costs without sacrificing performance.
• 3B parameter variant will power HP’s smart assistant ‘Xiaowei Hui’ in China, handling document generation and meeting summarization tasks (SCMP).
• Qwen team is also exploring self-improvement, aiming to evolve coding agents into autonomous, adaptive systems (Mint).

Why This Matters

• Enterprise-ready: Qwen3-Coder provides powerful long-context and coding-focused capabilities without the costs or restrictions of closed-source competitors.
• Benchmark leader: Outperforming GPT-4o and DeepSeek in rigorous mathematical and coding tests underscores Alibaba’s growing AI dominance.
• Productivity gains: Agentic capabilities and tool integration offer hands-free, intelligent handling of real-world software development tasks.
• Open-source edge: Companies can fully control infrastructure, scale usage, and customize the model—key advantages in regulated or security-sensitive industries.