all riversLarge ModelPublicis announced in a post today thatBaichuan IntelligenceLaunched one-stop big model commercializationSolutionBaichuan4-Turbo and Baichuan4-Air models and a link-wide domain enhancement toolchain are included to help enterprises realize cost-effective training.Private deploymentThe company has achieved multi-scenario availability rates of up to 96%.
It is reported that the program has the characteristics of "rich tools, fast response, significant effect, low cost", in the Baichuan4-Turbo, Baichuan4-Air On the basis of the algorithms of super-parametric dynamic search and adaptive rationing, and mixing and fine-tuning with enterprise private data, the availability rate of the two models in multiple scenarios can be greatly improved, and the average availability rate of the specialized segmentation tasks in finance, education, and healthcare scenarios is as high as 961 TP3T.
The main features of the two attached models are as follows:
Baichuan4-Turbo - Explore complex scenarios:
- The core capabilities of text generation, knowledge quiz, multi-language processing, data sub-clustering, etc. have been significantly improved, of which the information summary summarization capability has been substantially improved by 50%;
- Only 2 cards of 4090 arithmetic power are required for deployment;
- The inference cost is only 15% of Baichuan 4;
- Compared with Baichuan 4, the speed of Token has increased by 51% and the flow rate of Token has increased by 73%;
Baichuan4-Air - Proven scenarios for larger scale traffic:
- The effect is basically the same as Baichuan 4;
- The inference cost is only 1% of Baichuan 4;
- Millions of Token for only $0.98;
- Compared to Baichuan 4, the speed of the first Token has increased by 77% and the flow rate of Token has increased by 93%;
The official said that under the same training data, Baichuan4-Air not only has higher time efficiency, but also has a much better performance than MoE models with GPT4-style and Mixtral-style structure.
Currently, the program can be efficiently adaptedNVIDIA 4090 / A/H series, Huawei Rise, Cambrian, Qualcomm, MTK, TENAA, etc.A variety of mainstream chips.