Quen 1.5, the latest iteration of Alibaba’s AI model series, brings significant advancements in model quality and developer experience. Released just before the Chinese New Year, Quen 1.5 is an open-source model that competes closely with GPT-4 in performance and tool usage. It surpasses models like Claude 2.1 and GPT-3.5 Turbo in various evaluations, including MT Bench and Alpaca.
Quen 1.5 comes in six different sizes, ranging from 0.5 billion to 72 billion parameters, with each model excelling in different areas. The 0.5 billion model, for example, is exceptional at function calling, achieving a 77% pass rate on the Fireworks AI function calling evaluation dataset.
The video demonstrates Quen 0.5’s capabilities in function calling, showcasing how it can call external functions or APIs based on user input. This includes suggesting movies related to specific themes and compiling the best sources to fulfill user queries.
Quen 1.5’s performance is evaluated across various capabilities, including language understanding, coding, reasoning, and human preference alignment. It performs strongly in traditional benchmarks like MMLU, CEval, and HumanEval, and even surpasses GPT-4 in some Chinese language tasks.
The model also excels in aligning with human preferences, using techniques like Direct Policy Optimization (DPO) and Proximal Policy Optimization (PPO). It performs well in MT Bench and Alpaca evaluations, closely matching GPT-4 Turbo’s performance.
For those interested in using Quen 1.5, the video provides an installation guide using LM Studio. Users can easily download and run the model locally, selecting from various sizes and quantized versions like Int4, Int8, GPTQ, AWQ, and GGUF.
Overall, Quen 1.5 represents a significant leap in open-source AI models, offering robust performance and versatility across different tasks and languages. The video encourages viewers to explore Quen 1.5 and highlights its potential benefits for developers and AI enthusiasts.