Qwen: Qwen3 Coder Next

Qwen3-Coder-Next is an advanced open-weight causal language model specifically engineered for coding agents and streamlining local development workflows. Utilizing an innovative sparse Mixture-of-Experts (MoE) design, it boasts 80 billion total parameters while activating only 3 billion per token. This efficiency allows it to achieve performance levels akin to models requiring 10 to 20 times more active computation, making it exceptionally well-suited for cost-sensitive, always-on agent deployments. The model is meticulously trained with a strong agentic focus, ensuring reliable performance across long-horizon coding tasks, intricate tool usage, and robust recovery from execution failures. With a native 256k context window, Qwen3 Coder Next integrates seamlessly into real-world CLI and IDE environments, adapting effectively to common agent scaffolds used by modern coding tools. It operates exclusively in a non-thinking mode, simplifying integration for production coding agents by not emitting <think> blocks. Key capabilities include JSON mode, code generation, streaming, function calling, long context handling, and structured output. It features a 262K token context window and a max output of 65K tokens. Pricing is competitive at $0.12/$0.75 per 1M input/output tokens, available on the STARTER access tier.