Qwen1.5-7B-Chat

Qwen1.5 is a series of open-source language models developed to enhance model quality and developer experience, offering base and chat models ranging from 0.5 billion to 110 billion parameters. The models support a context length of up to 32,768 tokens and are available in quantized formats.

Key features:
- Model sizes: 0.5B, 1.8B, 4B, 7B, 14B, 32B, 72B, and 110B parameters
- Quantized models: Int4, Int8 GPTQ, AWQ, and GGUF formats
- Strong performance on benchmarks like MMLU, C-Eval, GSM8K, and HumanEval
- Human preference alignment using Direct Policy Optimization (DPO) and Proximal Policy Optimization (PPO)
- Multilingual support for 12 languages
- Long context support up to 32K tokens
- Retrieval-Augmented Generation (RAG) capabilities
- Tool use and function calling support

How it works:
- Models can be loaded using Hugging Face transformers (version 4.37.0 and above)
- Chat functionality utilizes `model.generate()` with chat templates
- Quantized models support low-resource deployment scenarios
- Local deployment is possible using frameworks like llama.cpp and Ollama

Integrations:
- Hugging Face Transformers for model loading and usage
- vLLM, SGLang, AutoAWQ, AutoGPTQ, Axolotl, and LLaMA-Factory for deployment, quantization, and finetuning
- llama.cpp and Ollama for local inference
- DashScope and together.ai for API services

Use of AI:
Qwen1.5 leverages generative artificial intelligence to enhance language understanding, reasoning, and multilingual capabilities. The models are aligned with human preferences using advanced techniques like DPO and PPO.

AI foundation model:
The document does not specify the AI foundation model used for Qwen1.5.

How to access:
- Web app: Accessible via platforms like Hugging Face
- API: Available through DashScope and together.ai
- Local deployment: Supported by frameworks like llama.cpp and Ollama

Qwen1.5 is suitable for researchers conducting advanced language model research, developers integrating language models into applications, and businesses requiring robust language understanding and generation capabilities. The models were launched in February 2024.