DeepSeek, a prominent Chinese AI enterprise, has unveiled a significant upgrade to its large language model, V3.1. This latest iteration embraces the UE8M0 data type, meticulously optimized to seamlessly integrate with the latest domestic chips. The upgrade merges the functionalities of both "thinking" and "non-thinking" models, empowering users to switch between modes effortlessly via chat templates. Furthermore, the context window capacity has been expanded substantially, from 65,536 Tokens to 131,072 Tokens.

In terms of performance, V3.1's non-thinking mode has demonstrated remarkable improvements across multiple key metrics, with a notable enhancement in tool invocation capabilities. It has achieved an impressive score of 30 in the Browsecomp test, surpassing its predecessor by a wide margin. Additionally, the new version effectively mitigates service costs by minimizing the usage of thinking Tokens.

DeepSeek has also made model weights accessible to developers, fostering further customization and optimization. This move underscores the company's commitment to fostering an ecosystem where innovation and flexibility thrive.