Qwen 3 Coder 480B Turbo

• 🧠 Mixture-of-Experts coder: 480B total, 35B active parameters
• ⚡ Turbo, FP8-quantized build tuned for faster code inference
• 📏 256K native context, extendable toward 1M via extrapolation
• 🔧 Agentic coding with a purpose-built function-call format
• 💬 Instruct, non-thinking model — no reasoning-trace blocks
• 🌐 Works with Qwen Code, CLINE, and similar agent tools
• 🏢 Built by Alibaba's Qwen team (Alibaba Cloud)
• 🎯 Capabilities here include function calling and web search

Alibaba

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Qwen 3 Coder 480B Turbo is a code-optimized large language model from Alibaba's Qwen team, served as a Turbo, FP8-quantized variant of the Qwen3-Coder-480B-A35B-Instruct base. The underlying model is a Mixture-of-Experts design with 480 billion total parameters and 35 billion active per inference, which the Qwen team frames as delivering high performance at lower compute cost than dense models of comparable scale. It supports a 256K-token context natively, with extrapolation methods reaching up to roughly 1M tokens.

The "Turbo" designation reflects an inference-optimized deployment: FP8 weights and provider-side serving aimed at faster, cheaper code workloads, which is the focus of this catalog entry. Functionally, it is an instruct, non-thinking model — it does not emit separate reasoning-trace blocks — and ships with a specially designed function-call format for agentic coding across tools like Qwen Code and CLINE.

Within Venice's broader Qwen lineup, it sits alongside general-purpose siblings such as Qwen 3 235B A22B Instruct 2507 and the efficiency-focused Qwen 3 Next 80b, but this checkpoint is specialized purely for coding and agentic tool use. As deployed here it adds function-calling and web-search capabilities for developer workflows.