Qwen1.5-110B release notes / what's new

Qwen has released Qwen1.5-110B, the first model in the Qwen1.5 series to exceed 100 billion parameters. This release demonstrates that increasing model size continues to provide performance gains in base capabilities and chat-based interactions, positioning the model as a competitive alternative to other large-scale open-source models like Llama-3-70B.

Model Architecture and Technical Specifications

Qwen1.5-110B utilizes a Transformer decoder architecture similar to other models in the Qwen1.5 series. Key technical specifications include:

  • Grouped Query Attention (GQA): The model incorporates GQA to ensure efficiency during model serving.
  • Context Window: It supports a context length of 32K tokens.
  • Multilingual Support: The model maintains multilingual capabilities, supporting a wide range of languages including English, Chinese, French, Spanish, German, Russian, Korean, Japanese, Vietnamese, and Arabic.

Base Model Performance

Qwen1.5-110B is competitive with Meta-Llama3-70B and Mixtral-8x22B across several benchmarks. The Qwen team notes that performance improvements over the Qwen1.5-72B model were achieved by increasing the model size without drastically changing the pretraining and post-training recipes.

Benchmark Qwen1.5-110B Qwen1.5-72B Llama-3-70B Mixtral-8x22B
MMLU 80.4 77.5 79.5 77.8
TheoremQA 34.9 29.3 32.0 35.9
GPQA 35.9 36.3 36.4 34.3
Hellaswag 87.5 86.0 88.0 88.7
BBH 74.8 65.5 76.6 69.2
ARC-C 69.6 65.9 68.8 70.7
GSM8K 85.4 79.5 79.2 78.6
MATH 49.6 34.1 41.0 41.7
HumanEval 52.4 41.5 45.7 45.1
MBPP 58.1 53.4 55.1 71.2

Chat Model Evaluation

In chat-specific evaluations, Qwen1.5-110B-Chat shows significant improvement over the Qwen1.5-72B-Chat model. It outperforms Llama-3-70B-Instruct in both MT-Bench and AlpacaEval 2.0.

Model MT-Bench (Avg. Score) AlpacaEval 2.0 (LC Win Rate)
Llama-3-70B-Instruct 8.85 34.40
Qwen1.5-72B-Chat 8.61 36.60
Qwen1.5-110B-Chat 8.88 43.90

Implications for Model Scaling

The release of Qwen1.5-110B indicates that there is still significant room for performance gains through model size scaling. While the Qwen team acknowledges the importance of data scaling as demonstrated by Llama-3, they intend to pursue a strategy of scaling both data and model size for future releases, including the upcoming Qwen2.

Sources