Qwen3-30B-A3B-Thinking-2507

Qwen3-30B-A3B-Thinking-2507 is an upgraded hybrid version of the Qwen3-30B-A3B model, specifically optimized for reasoning-only mode, significantly enhancing its reasoning capabilities. Built on a Mixture of Experts (MoE) architecture, the model has 30.5 billion total parameters, with only 3.3 billion activated per inference. Out of 128 experts, just 8 are activated per task, enabling dynamic adaptation to diverse query types. The model features 48 hidden layers and employs Group Query Attention (32 query heads and 4 key-value heads), ensuring efficient information processing while maintaining high-quality attention mechanisms. Architectural innovations also include native support for an extended context length of up to 262,144 tokens, making the model ideal for analyzing large documents, complex codebases, and performing multi-step reasoning.

The advanced reasoning mode enables Qwen3-30B-A3B-Thinking-2507 to achieve outstanding results on the AIME25 math benchmark (85.0), surpassing the closely sized proprietary model Gemini 2.5-Flash-Thinking (72.0). The model also excels in agent-like use cases, scoring 72.4 on the BFCL-v3 benchmark, making it an excellent choice for integration with external tools to automate complex workflows. Notably, for highly complex tasks, developers are recommended to use an output length of up to 81,920 tokens, allowing the model to fully leverage its potential in step-by-step reasoning. For routine tasks, a standard output length of 32,768 tokens is sufficient.

In summary, Qwen3-30B-A3B-Thinking-2507 is a versatile solution for large industrial enterprises, research centers, and educational institutions—where high-level analytical reasoning is required, and a mid-sized yet powerful model is preferred.

Announce Date: 29.07.2025
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: 4.51.0
License: Apache 2.0

Use our pre-built public endpoints to test inference and explore Qwen3-30B-A3B-Thinking-2507 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.