Qwen2.5-3B

Qwen2.5-3B features 3 billion parameters, 36 layers, and a 16/2 attention head architecture, delivering a significant performance leap while maintaining reasonable resource requirements. The model supports a 32K-token context window and up to 8K-token generation, enabling it to handle moderately complex tasks with extended contexts.

The uniqueness of Qwen2.5-3B lies in its return to the product line after being absent from the Qwen2 series, effectively filling the crucial gap between 1.5B and 7B models. This size proves particularly valuable for resource-constrained scenarios where the 7B version might be excessive, yet higher performance than the 1.5B variant is required. The model demonstrates substantially improved capabilities in understanding complex instructions, multi-step reasoning, and working with structured data.

Notably, this model is distributed under the Qwen Research License, which may impose certain restrictions on commercial use. However, Qwen2.5-3B is ideally suited for research projects, prototyping, and developing specialized solutions where licensing flexibility for research purposes is essential. The model performs exceptionally well in data analysis tasks, technical documentation processing, educational applications, and serves as an excellent base for creating domain-specific models through fine-tuning.

Announce Date: 19.09.2024
Parameters: 3B
Context: 32K
Attention Type: Full Attention
VRAM requirements: 2.5 GB using 4 bits quantization
Developer: Alibaba
Transformers Version: 4.43.1
License: qwen

Use our pre-built public endpoints to test inference and explore Qwen2.5-3B 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.