Alibaba's Qwen team unveils Qwen3-Coder-Next, an 80-billion-parameter model built on a sparse Mixture-of-Experts architecture designed for agentic coding and local development, leveraging a hybrid Gated DeltaNet and Gated Attention stack for long-context processing.

Specialized Expert Models for Web Development and UX are trained and distilled into the main 80B/3B MoE model to retain domain expertise while staying lightweight.

Architecturally, it combines Gated DeltaNet, Gated Attention, and Mixture-of-Experts in a 48-layer stack with a 2048 hidden dimension, featuring 512 MoE experts and a per-token mix of 10 experts plus 1 shared expert.

Deployment is supported through SGLang and vLLM with OpenAI-compatible endpoints, including local GGUF quantizations (4-bit around 46 GB RAM, 8-bit around 85 GB) and a token context up to 262,144, with practical defaults nearer 32,768 for smaller machines.

MegaFlow orchestrates a three-stage workflow—agent rollout, evaluation, and post-processing—with closed-loop feedback enabling real-time learning from environments.

Agentic training encompasses large-scale executable task synthesis, environment interactions, and reinforcement learning, totaling roughly 800,000 verifiable tasks to support long-horizon planning, tool usage, and recovery from failures.

Key takeaways highlight MoE efficiency, long-horizon coding capabilities, agentic training, strong benchmark performance, and practical local deployment under Apache-2.0.

Supporting resources include a technical report, code repositories, model weights, and deployment guidance across multiple platforms.

Core breakthroughs enable a 262,144-token context window and linear-time processing by fusing Gated DeltaNet with Gated Attention in a scalable long-sequence framework.

The model runs with only about 3 billion active parameters per token, delivering high throughput and lower inference costs while aiming to match the performance of larger active models.

Qwen3-Coder-Next maintains an 80B parameter size but activates a small subset per token, targeting efficient long coding sessions and agent workflows.

The agentic pipeline is implemented on Alibaba Cloud Kubernetes, sourcing real-world bug fixes and executable environments to generate verifiable tasks.