China’s DeepSeek adds expert chatbot mode ahead of much-awaited V4 release

DeepSeek has introduced new “instant” and “expert” modes to its chatbot, the most significant update to its user interface since the Hangzhou-based start-up gained global recognition. The changes come ahead of the much-anticipated release of DeepSeek’s next-generation flagship model V4 this month, more than a year after its R1 version made it a household name. On Tuesday, the company added the two modes to its website and mobile app, giving users the choice of settings. Instant mode was designed for quick replies during “daily conversations”, while expert mode was intended for those with “complex problems” though it also warned that the service could be “busy at peak times”. Artificial intelligence chatbots including Google’s Gemini and ByteDance’s Doubao also allow users to choose different usage modes to suit their needs. Expert modes generally deliver higher-quality responses, though they are more computationally demanding. The South China Morning Post on Wednesday tested the instant and expert modes. Asked to create a JavaScript animation of falling letters with realistic physics, the instant mode produced code in over a minute that followed the instruction, while the expert mode took around 40 seconds to generate code that failed to meet the instruction. The expert mode has also removed the option to upload files – a crucial feature for productivity tasks such as data processing – while the instant mode has retained file attachment but only in text form. Leading AI models from the likes of OpenAI and Anthropic are multimodal, meaning that they can process image and video in addition to text. The limitations of the new features have invited speculation online that the chatbot’s expert mode has yet to be integrated with DeepSeek’s next-generation V4 model, which is expected to deliver major improvements, possibly including multimodal capabilities. DeepSeek did not respond to a request for comment. The start-up has been the subject of much attention in recent months, with expectations it would release its next major model during the Chinese New Year period, repeating its playbook from last year. However, the huge jump in size of the model to over 1 trillion parameters – the mathematical variables that encode a model’s “intelligence” – had delayed its release, SCMP previously reported. Interest also grew after DeepSeek’s chatbot suffered a prolonged outage in late March, disrupting service for hundreds of millions of users. Despite the setbacks, the company is still seen as the Chinese foundational AI model developer best placed to narrow the gap with US rivals. On Wednesday, US start-up Anthropic, widely considered among AI researchers as the leading foundational AI model developer in the world, announced a new model called Claude Mythos Preview, which it said was too powerful for public release. The start-up, which recently engaged in a high-profile battle with the Pentagon over use of its Claude models, said the model would be used internally by a consortium of leading US tech giants – including Apple, Nvidia and Microsoft – to patch security vulnerabilities in critical software systems.