Editorial illustration for Qwen3-Coder-Next: 10× throughput beats Claude‑Opus‑4.5 on SecCodeBench
Qwen3-Coder-Next: Open Source AI Beats Top Coding Models
Qwen3-Coder-Next: 10× throughput beats Claude‑Opus‑4.5 on SecCodeBench
A three billion parameter model just beat Claude Opus 4.5 on a security test. It wasn't close. Alibaba's new Qwen3-Coder-Next scored 61.2% on SecCodeBench, a benchmark that measures whether a model can spot and fix vulnerabilities without being told to look for them. It did this with ten times the throughput, processing a 262,000-token code repository in seconds.
The model is a sparse expert. Only three billion of its parameters are active at any time. This is the core of the shock.
It is fast, small enough to run in a Docker container, and it beat one of the best closed models. The result suggests raw size is becoming a liability.
It also posted a competitive 56.32% func-sec@1 score on CWEval, outperforming DeepSeek-V3.2 and GLM-4.7. The takeaway is blunt. After training on 800,000 agentic tasks, the model learned to anticipate security flaws. It balances function and safety by default.
The core technical breakthrough behind Qwen3-Coder-Next is a hybrid architecture designed specifically to circumvent the quadratic scaling issues that plague traditional Transformers.
The competition is no longer about who has the biggest brain. It is about who can use a smaller one the fastest. A model that scans an entire codebase before a giant model finishes its preamble is a different kind of tool.
Practicality has changed the definition of power. This sparse, open expert isn't just challenging the giants. It is making them look slow.
Related Reading
Browse all Anthropic news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic Anthropic says Claude writes >90% of its code; AI pause button urged
LLMs & Generative AI
Related Topic Claude Mythos highlights EU AI safety gaps, says researcher Caroli
Policy & Regulation
From Archives Vercel revamps v0 with sandbox runtime to deliver production‑ready AI code
From Archives AI agents launch dedicated social network as GitLab showcases roadmap
From Archives Anthropic teams with Allen Institute and HHMI to boost transparent scientific AI
Research & Benchmarks
From Archives Apple adds Claude and Codex to Xcode 26.3, enabling auto‑entitlement for APIs
LLMs & Generative AI
Common Questions Answered
How does Qwen3-Coder-Next perform on security vulnerability detection and repair?
The model demonstrated exceptional performance on SecCodeBench, achieving a 61.2% success rate in code generation scenarios that involve vulnerability repair. Notably, it maintained high scores even without explicit security hints, suggesting an inherent understanding of potential security pitfalls developed during its 800k-task agentic training phase.
What makes Qwen3-Coder-Next unique in terms of performance and architecture?
Qwen3-Coder-Next features an ultra-sparse architecture that claims to deliver 10× throughput compared to systems like Claude-Opus-4.5. The model is designed for repository-wide tasks and represents Alibaba's continued effort to create open-source AI models that can compete with or exceed proprietary systems.
What is significant about the model's multilingual security evaluations?
The model demonstrated robust performance across multilingual security evaluations, indicating its ability to detect and repair vulnerabilities across different programming languages. This multilingual capability suggests a sophisticated understanding of code security that transcends language-specific boundaries.
