Parameters
14B
Context Length
131.072K
Modality
Text
Architecture
Dense
License
Apache 2.0
Release Date
29 Apr 2025
Knowledge Cutoff
-
Attention Structure
Grouped-Query Attention
Hidden Dimension Size
-
Number of Layers
48
Attention Heads
80
Key-Value Heads
8
Activation Function
-
Normalization
Layer Normalization
Position Embedding
ROPE
VRAM requirements for different quantization methods and context sizes
Qwen3-14B is a causal language model developed by the Qwen team at Alibaba Cloud, belonging to the Qwen3 series. This model is engineered with a dense architecture, encompassing 14.8 billion parameters. A core aspect of its design is the capacity for dynamic mode switching between a "thinking" mode for intricate analytical tasks and a "non-thinking" mode for efficient general-purpose dialogue. This dual operational capability aims to optimize performance and utility across a broad spectrum of natural language processing applications.
From an architectural standpoint, Qwen3-14B incorporates a Grouped Query Attention (GQA) mechanism, configured with 40 query heads and 8 key/value heads, which contributes to its computational efficiency. The model is structured with 40 layers. It supports a native context length of 32,768 tokens, which can be expanded to 131,072 tokens through the application of the YaRN (Yet another RoPE N) technique for Rotary Position Embeddings. Further architectural refinements include the implementation of qk layernorm, which is integrated across all Qwen3 models to enhance training stability and overall performance.
In terms of its operational characteristics, the thinking mode of Qwen3-14B demonstrates enhanced reasoning capabilities, particularly in domains such as mathematics, code generation, and complex logical inference. Conversely, the non-thinking mode is optimized for tasks requiring general dialogue, instruction following, and creative content generation. The model supports over 100 languages and dialects, showcasing robust multilingual processing capabilities. Its design also facilitates integration with external tools, endowing it with agentic functionalities for addressing complex, multi-step problems. These features position Qwen3-14B as a versatile asset for applications ranging from advanced AI assistants requiring analytical depth to interactive conversational systems.
The Alibaba Qwen 3 model family comprises dense and Mixture-of-Experts (MoE) architectures, with parameter counts from 0.6B to 235B. Key innovations include a hybrid reasoning system, offering 'thinking' and 'non-thinking' modes for adaptive processing, and support for extensive context windows, enhancing efficiency and scalability.
Ranking is for Local LLMs.
Rank
#11
Benchmark | Score | Rank |
---|---|---|
Data Analysis LiveBench Data Analysis | 0.68 | ⭐ 4 |
Reasoning LiveBench Reasoning | 0.74 | 7 |
Mathematics LiveBench Mathematics | 0.73 | 9 |
Coding LiveBench Coding | 0.58 | 12 |
Overall Rank
#11
Coding Rank
#20
Full Calculator
Choose the quantization method for model weights
Context Size: 1,024 tokens