Qwen3-Coder-30B-A3B-Instruct

Qwen3-Coder-30B-A3B-Instruct is an outstanding example of a high-quality large language model with advanced specialization in programming. This Mixture-of-Experts model has 30.5 billion total parameters, of which only 3.3 billion are activated per token, and out of 128 experts, only 8 are activated per token. The model comprises 48 hidden layers with grouped query attention (32 heads for Q and 4 for KV), delivering exceptional processing efficiency with minimal computational resource consumption. Native support for a 262,144-token context window—expandable up to 1 million tokens via Yarn—makes the model ideal for working with large code repositories within complex projects.

The key unique feature of Qwen3-Coder-30B-A3B-Instruct lies in its superior agent capabilities. The model does not merely generate code; it autonomously interacts with development tools, executes multi-step programming tasks, and is capable of solving complex problems without human intervention. On the LiveCodeBench v6 benchmark, the model achieves an impressive 66.0%, significantly outperforming the base version Qwen3-30B-A3B (57.4%). In AIME25 tasks (advanced mathematics for programming), it demonstrates 85.0% accuracy, surpassing Gemini-2.5-Flash-Thinking (72.0%) and confidently competing with much larger models. The model outperforms DeepSeek V3 on most coding tasks and delivers agent workflow performance comparable to Claude Sonnet 4, a remarkable achievement for an open-source solution.

Qwen3-Coder-30B-A3B-Instruct unlocks entirely new possibilities in software development. The model is integrated with popular agent-based programming platforms, including Qwen Code, CLINE, Roo Code, and Kilo Code, offering a unified function-calling format for seamless operation within CI/CD pipelines. Support for 358 programming languages makes it a universal reference tool for developers. The model particularly excels in repository-scale understanding scenarios, where it can analyze and modify massive codebases, automatically refactor legacy code, and create complex full-stack applications with minimal developer intervention.

Announce Date: 22.07.2025
Parameters: 30.5B
Experts: 128
Activated: 3.3B
Context: 263K
Attention Type: Full or Sliding Window Attention
VRAM requirements: 37.1 GB using 4 bits quantization
Developer: Alibaba
Transformers Version: 4.52.3
License: Apache 2.0

Use our pre-built public endpoints to test inference and explore Qwen3-Coder-30B-A3B-Instruct capabilities.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.