Qwen3 Coder 30B A3B Instruct (NovitaAI)

Qwen3-Coder-30B-A3B-Instruct is an open-weights coding-focused large language model published by the Qwen team and distributed under the Apache-2.0 license, positioned within the broader Qwen3 family of causal language models. The variant documented on ModelScope lists roughly 30.5 billion total parameters with a Mixture-of-Experts design that activates around 3.3 billion parameters per pass, paired with 48 layers and grouped-query attention across 32 heads, giving it a sparse-but-capable profile aimed at programming workloads rather than dense general-purpose chat. Because the weights are openly released, developers can self-host, fine-tune, or inspect the model directly, which is central to its intended use as a foundation for code agents and developer tools.

The model is framed by its publisher around three practical strengths: strong agentic coding ability, agentic browser-use behavior, and very long context handling, with native support advertised at hundreds of thousands of tokens and extrapolation up to one million tokens via YaRN for repository-scale understanding. Its training combines pretraining and post-training stages, and the official notes highlight a specialized function-call format that makes it suitable for integration with coding assistants and agent frameworks such as Qwen Code and Cline. For practitioners, the practical fit is a self-hostable coding model that can ingest large codebases, drive multi-step programming agents, and expose structured tool calls, while remaining light enough at inference time thanks to its sparse activation pattern.