Qwen3-Coder-30B-A3B-Instruct is a 30.5B parameter Mixture-of-Experts (MoE) model with 128 experts (8 active per forward pass), designed for advanced code generation, repository-scale understanding, and agentic tool use. Built on the Qwen3 architecture, it supports a native context length of 256K tokens (extendable to 1M with Yarn) and performs strongly in tasks involving function calls, browser use, and structured code completion.
This model is optimized for instruction-following without “thinking mode”, and integrates well with OpenAI-compatible tool-use formats.
Evaluations
13
across 13 benchmarks
Latency
1.3s
Context length
262k
tokens
Cost
$0.06 · $0.25
input · output per 1M tokens
Key takeaways
Qwen3 Coder 30B A3B Instruct is a 30.5B Mixture-of-Experts (MoE) model built on the Qwen3 architecture, excelling in code generation, repository-scale understanding, and agentic tool use (function calls, browser use, structured code completion).
It supports a native context length of 256K tokens, extendable to 1M with Yarn, and is optimized for instruction-following without “thinking mode”, integrating well with OpenAI-compatible tool-use formats.
The model demonstrates exceptional proficiency in mathematical reasoning, achieving 94.16% accuracy on Grade School Math 8K, consistently providing accurate numerical answers and executing multi-step mathematical problems effectively.
It struggles significantly with complex arithmetic, numerical precision errors, misinterpreting ambiguous problem constraints, and lacks external/up-to-date specialized knowledge, particularly evident in its 3.37% accuracy on 'Humanity's Last Exam'.
On 'Humanity's Last Exam', the model exhibited strong analytical and structured reasoning but had issues with numerical accuracy in multi-step arithmetic calculations and with highly specialized or less common factual knowledge, leading to a low accuracy of 3.37% on this benchmark.
