Qwen3-4B-Instruct-2507

Qwen3-4B-Instruct-2507 is a revolutionary model built on an innovative architecture with 4.02 billion parameters (including embeddings), 36 transformer hidden layers, and Group Query Attention (GQA) using 32 attention heads for queries and 8 for keys and values—providing an optimal balance between performance and memory efficiency. The model is optimized from the hybrid Qwen3-4B base to operate exclusively in non-thinking mode, completely eliminating the generation of <think></think> blocks, thereby maximizing query processing speed. Native support for a context length of 262,144 tokens enables efficient handling of large documents, extended conversations, and complex multi-step tasks without degradation in information processing quality.

Architectural innovations include an advanced user-preference alignment system, delivering more relevant and useful responses, along with significant improvements in multilingual content processing.

The model demonstrates outstanding results on key benchmarks, outperforming the proprietary GPT-4.1-nano across all major metrics: MMLU-Pro (69.6 vs 62.8), GPQA (62.0 vs 50.3), and particularly impressive scores on ZebraLogic (80.2 vs 14.8) and creative content generation, where it achieves 83.5 (vs 72.7). The model excels in instruction-following tasks, achieving 83.4% on IFEval and 43.4 on Arena-Hard v2. It also performs exceptionally well in agent-based tasks and tool usage, showing strong results on the BFCL-v3 (61.9) and TAU benchmark suites, making it ideal for integration into automated systems.

Qwen3-4B-Instruct-2507 is highly suitable for business process automation, including customer service via intelligent chatbots, document processing and analysis, report generation, and personalized recommendations. It is effective in creating and localizing SEO-optimized marketing content, product descriptions, social media posts, and more. Thanks to seamless API integration, the model can be deployed for automation within CRM and ERP systems, as well as for any tasks requiring intelligent routing and fast, real-time query processing.

Announce Date: 07.08.2025
Parameters: 4.02B
Context: 263K
Attention Type: Full or Sliding Window Attention
VRAM requirements: 36.2 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 Qwen3-4B-Instruct-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.