GPT4 Turbo is one of the best LLMs of our time. The latest LLM from Chinese company SenseTime, SenseNova 5.0, beats it across the board.
The new iteration of SenseNova, SenseNova 5.0 is a “full-stack large model matrix” that was unveiled a couple of days ago. It’s multimodal and has excellent reasoning, coding, and logical performance that’s ideal for finance and data analysis, among other domains. This is pretty much the key USP of OpenAI’s GPT4.
It’s a 600B parameter model with a whopping 200k context window, being trained on 10TB+ of largely synthetic data. CEO Xu Li announced the new model among other highlights at the company’s Tech Day event.
According to the China Daily report, “At the application level, SenseNova 5.0 supports high-resolution long-image parsing, understanding, interactive text-to-image generation, complex cross-document knowledge extraction, summarization, question-answering display, and rich multimodal interactions.“
The tagline for the model on the official page of the SenseTime website is “Linguistic, creative and scientific capabilities greatly improved; multimodal interactions added.“
The company is calling it industry-leading. And it certainly is.
In a comparison published by OpenCompass, it’s better than GPT4 Turbo in 14 benchmarks. That’s pretty much all the benchmarks that AI companies use to give their models a technical analysis. Most notably, the SenseNova 5.0 is 121% better than GPT4 in 0-shot MATH, 16.6% better in HellaSwag, 33.5% better in CMMLU, and 30.5% better in MBPP. In HumanEval, the model only has a 1.67% improvement.
The chart also compared it to Llama and the gains were just insane, as expected. The model needs to be open to the rest of the world for a while before we can see more comparisons, human scores, and how it stacks up against Gemini, though there’s all the reason to believe that it’s better than Gemini Ultra as well.
