ApX logoApX logo

Qwen3-8B

Parameters

8B

Context Length

131K

Modality

Text

Architecture

Dense

License

Apache 2.0

Release Date

29 Apr 2025

Knowledge Cutoff

-

System Requirements

VRAM requirements for different quantization methods and context sizes

1,024 tokens

18.48 GB VRAM

Consumer

1x RTX 4090

24GB VRAM

Datacenter

1x NVIDIA A100

80GB VRAM

Apple Silicon

1x Apple M3 Max

128GB VRAM

131,072 tokens

40.85 GB VRAM

Consumer

2x RTX 4090

24GB VRAM

Datacenter

1x NVIDIA A100

80GB VRAM

Apple Silicon

1x Apple M3 Max

128GB VRAM

Architecture Diagram

Input TokensToken EmbeddingPosition: RoPEHidden: 4.1k · Context: 131K · Vocab: 151.9kx 40 layersLayerNormPre-AttentionGrouped-Query Attention64Q / 8KV headsHead dim: 128+LayerNormPre-FFNFeed-Forward NetworkSwiGLUIntermediate: 12.3k+Final LayerNormOutput Logits

Evaluation Benchmarks

Rank

#49

BenchmarkScoreRank

General Knowledge

MMLU

0.852

14

Rankings

Overall Rank

#49

Coding Rank

-

About Qwen3-8B

Qwen3-8B is a dense causal language model developed by Alibaba, part of the broader Qwen3 series. It consists of approximately 8.2 billion parameters and is engineered for efficient performance across a spectrum of natural language processing tasks. A distinctive feature within the Qwen3 family is the integration of a "thinking" mode for complex logical reasoning, mathematics, and coding, alongside a "non-thinking" mode optimized for general-purpose dialogue. This design facilitates dynamic adaptation of the model's operational characteristics based on task demands without requiring a switch between distinct models.

The architectural foundation of Qwen3-8B is the decoder-only transformer, incorporating refinements such as qk layernorm for enhanced stability and leveraging Grouped Query Attention (GQA) to optimize inference speed and memory utilization by sharing Key/Value heads among multiple Query heads. Its training regimen is a three-stage process, involving extensive pre-training on over 36 trillion tokens across 119 languages to build broad language proficiency and general knowledge. This initial stage (S1) is followed by specific optimization for reasoning skills in a second stage (S2) by increasing the proportion of STEM, coding, and reasoning data, and long-context comprehension in a third stage by extending training sequence lengths up to 32,768 tokens natively. The context length can be further extended to 131,072 tokens via the YaRN method.

Qwen3-8B exhibits enhanced reasoning capabilities and superior human preference alignment, making it effective for applications requiring creative writing, role-playing, multi-turn dialogues, and precise instruction following. Furthermore, it includes agent capabilities, supporting integration with external tools for complex agent-based tasks. The model's comprehensive multilingual support extends to over 100 languages and dialects, facilitating multilingual instruction following and translation.

Technical Specifications

Attention

Attention Structure

Grouped-Query Attention

Attention Heads

64

Key-Value Heads

8

Attention Head Dimension

128

Position Embedding

ROPE

RoPE Theta

1,000,000

Sliding Window Attention

No

Sliding Window Size

-

Sliding Window Ratio

-

Linear Attention

-

Linear Attention Ratio

-

Normalization

Layer Normalization

Activation Function

SwigLU

Dimensions

Hidden Dimension Size

4,096

Number of Layers

40

FFN Intermediate Size (Dense)

12,288

Multi-Token Prediction Heads

-

Tokenizer

Vocabulary Size

151,936

About Qwen 3

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.


Other Qwen 3 Models