Community

Blog Events Webinars Tutorials Forum
  1. Blog
  2. Events
  3. Webinars
  4. Tutorials
  5. Forum
Create Account
×
Community Blog Qwen-Coder-Qoder: Customizing a Fast-Evolving Frontier Model for Real Software

Qwen-Coder-Qoder: Customizing a Fast-Evolving Frontier Model for Real Software

Inside Qwen-Coder-Qoder: The RL-optimized model behind Qoder's agent architecture

Learn More about Qoder

Explore Qoder for Enterprise

Introduction

Today, we are pleased to introduce Qwen-Coder-Qoder, a customized model tailored to elevate the end-to-end agentic coding experience on Qoder.

Built upon the Qwen-Coder foundation, Qwen-Coder-Qoder has been optimized with large-scale reinforcement learning to align tightly with Qoder's scenarios, tools, and agent architecture. On Qoder Bench — our benchmark for real-world software engineering tasks — it surpasses Cursor Composer-1 in task resolution performance. The gains are particularly notable on Windows, where terminal command accuracy is improved by up to 50%.

1

At the same time, Qwen-Coder-Qoder has delivered tangible, data-backed improvements to the Qoder user experience. With rapid model iterations, we have observed meaningful gains in production over the past few weeks: code retention has increased by 3.85%, tool error rates have dropped by 61.5%, and token consumption has decreased by 14.5%. These metrics now position Qoder alongside the world's top-tier models.

2

Beyond superior performance metrics, Qwen-Coder-Qoder reflects the 'taste' and 'mindset' of a senior software engineer. We believe a great AI coding partner shouldn't just solve problems—it should solve them elegantly and masterfully.

  • Adheres to best engineering practices: Many general models optimize solely for task resolution and may take shortcuts that bypass established engineering conventions. In contrast, Qwen-Coder-Qoder is trained to follow rigorous software engineering principles, maintain consistent project code style, and ensure production-ready outputs.
  • Holistic Repository Understanding: By leveraging Qoder's unique context systems — including code graphs, project memory, and Repo Wiki — Qwen-Coder-Qoder understands the project from a global perspective and uses the right tools to complete tasks with precision.
  • High-Efficiency Parallelism: The model recognizes tasks that don't depend on each other and runs them parallelly — whether it's fetching code, planning tasks, or making multiple edits. This makes the entire workflow much faster.
  • Resilient Problem Solving: When faced with complex or stubborn issues, general models may abandon the task after limited attempts. Qwen-Coder-Qoder demonstrates a developer-level persistence: it refines its approach iteratively and stays engaged until the problem is resolved.

Our Vision: A "Model-Agent-Product" Flywheel for Co-evolving Intelligence

Qwen-Coder-Qoder is not an accident — it is the inevitable outcome of the intelligent evolution loop we've built around the Model-Agent-Product paradigm.

3

In the rapidly evolving landscape of AI coding, we've focused on building a self-evolving cycle: Model → Agent → Product (model as agent, agent as product, product reinforces model). This loop ensures that insights from real user interactions continuously inform and enhance our models' capabilities. At the core of this system, the model provides the foundation — we embed all of the capabilities required by the Qoder Agent directly into Qwen-Coder-Qoder, which powers task execution. On the product side, the Agent is central — every feature and workflow in Qoder revolves around it. With thousands of users engaging with the product daily, we capture real-world usage patterns and preferences, extract best software engineering practices, and convert them into reward signals that further strengthen our RL training.

This completes our flywheel of software engineering intelligence. Qwen-Coder-Qoder is a large-scale RL model trained on real-world product environments, real-world development tasks, and real-world rewards.

Under the Hood: How We Made It Happen

Achieving these results requires a robust, state-of-the-art training strategy built on three core elements:

A Real-World Qoder Agent as the Sandbox

We train the model to master the full stack of Qoder's Knowledge, Memory, Tools/MCP, and Context to solve real-world coding tasks. Unlike general-purpose models, our model is tightly aligned with the Qoder product. As the model continues to evolve, this synergy is unlocking massive value. To scale this, we've automated the setup of tens of thousands of real-world software environments. Using High-speed containerization, we can spin up and tear down these sandboxes instantly to power our reinforcement learning at massive scale.

4

Real-World Best Engineering Practices as Reward Signals

In agentic reinforcement learning, reward signals are critical to guiding the model toward desirable behaviors. We use several criterias to verify correctness including unit tests, CLI checks, and custom checklists to make sure the agent actually gets the job done. It's not just about getting a passing diff patch. We also enforce strict rules on how the code is written, ensuring the entire process follows professional engineering standards. we ensure the agent's output meets the same standards you'd expect from a senior engineer.

Reward hacking is an inherent challenge in reinforcement learning. For instance, if we reward the model for parallel tool use to boost speed, it might try to 'game the system' by scanning tons of irrelevant files just to rack up points. While the parallelism metrics look great, there's no real contribution to the final accuracy.

Solving Reward Hacking is essentially a battle of wits with the model. To tackle this, we built a 'Rewarder-Attacker' adversarial framework. We use an LLM as a reviewer to constantly stress-test and 'attack' our reward system, hunting for loopholes before training even begins. This setup has drastically improved both the iteration speed and the robustness of our reward design.

Large-Scale, High-Efficiency RL Training Framework

Qwen-Coder-Qoder is powered by ROLL, which is optimized to enable efficient RL training of MoE LLMs with hundreds of billions of parameters on clusters scaled to thousands of GPUs. In a typical RL loop, the rollout phase often consumes over 70% of the total time. To maximize end-to-end throughput, we optimized the system from two angles:

  • Optimizing the Rollout Phase: We implemented asynchronous scheduling to minimize idle time, Prefix/KV cache reuse to eliminate redundant compute, and redundant environment execution to mitigate long-tail latency.
  • Rollout-Training Co-design: We decoupled the two by relaxing on-policy constraints to allow cross-version sampling. By running training and rollout in parallel, we implemented dynamic resource yielding, ensuring that GPUs are surrendered to rollout during training wait times.

Together, these system-level optimizations delivered a 10x boost in throughput, significantly compressing our training cycles.

Future Prospects

The Qwen-Coder-Qoder we're releasing today is the first milestone of our "Model-Agent-Product" flywheel. In just a few short months, we've already witnessed how this loop can drastically elevate the end-to-end experience.

5

This is just the beginning. Doubling down on this trajectory, we will continue to evolve through weekly iterations, refining model efficacy and experience as we forge ahead toward an 'Agentic Coding Platform for Real Software'.

AI Artificial Intelligence Developers Qwen AI Coding AI Agent Agentic Coding Qoder Qwen-Coder-Qoder QODER CLI Qwen-Coder
0 0 0
Share on

Read previous post:

In Practice: QODER CLI × HARNESS ENGINEERING — Building a 7×24 Autonomous Feedback Processing System

Alibaba Cloud Community

1,462 posts | 503 followers

You may also like

Comments

Alibaba Cloud Community

1,462 posts | 503 followers

Related Products

More Posts by Alibaba Cloud Community

See All