00:00DeepSeek has unveiled two new versions of an experimental AI model that rivals offerings from OpenAI and Google.
00:08The update adds fresh capabilities, the startup says, will help with combining reasoning and executing certain actions autonomously.
00:15For more, let's bring in Asia Tech correspondent Annabelle Druler.
00:18So, Belle, how powerful is this AI model?
00:22Yeah, well, to your point there, they're saying that it's really comparable to what you're seeing from the latest,
00:27from the likes of alphabets, Gemini, for instance, also OpenAI as well.
00:32So, certainly at a benchmark, the DeepSeek says rivals the best technology coming out of the States.
00:37But just to take a step back and talk about exactly what's happened here.
00:41So, in September of this year, DeepSeek released the so-called V3.2 XP, or EXP, a nod essentially to its experimental status.
00:51Now we've actually got two new versions of that experimental model that have come to market.
00:57So, there's the V3.2, and the company is saying this one here, it has, again, strong capabilities.
01:05They're comparing it to ChatGPT5, for instance, from OpenAI on a range of different performance benchmarks.
01:11They're also saying it has some agentic capabilities inbuilt.
01:14So, agentic capabilities, you're adding a sort of human reasoning here, and they're saying that this is a model that can use tools like search engines, calculators, code executors as well.
01:24So, that's the first version.
01:26The second version they're calling V3.2 Special.
01:30And this is a version that they're saying actually has an ability to focus on mathematical computations, more long-term, more long-thinking capabilities as well.
01:40And they're saying that the goal here is really to push the inference capabilities on open-source models to their limits, explore the capabilities of models as well.
01:49But they're saying that that one matches the performance of the latest Gemini 3 Pro.
01:54So, again, going back to your point, yes, these are essentially a Chinese player going that open-source approach, doing more with less, but able, they're saying, to meet the best from the States.
Be the first to comment