Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More. Transformer-based large language models (LLMs) are the foundation of the modern generative AI landscape.Transformers aren’t the only way to do gen AI, though. Over the course of the last year, Mamba, an approach that uses Structured State Space Models (SSM), has also picked up adoption as an alternative approach from multiple vendors, including AI21 and AI silicon giant Nvidia. Nvidia first discussed the concept of Mamba-powered models in 2024 when it initially released the MambaVision research and some early models. This week, Nvidia is expanding on its initial effort with a series of updated MambaVision models available on Hugging Face.MambaVision, as the name implies, is a Mamba-based model family for computer vision and image recognition tasks. The promise of MambaVision for enterprise is that it could improve the efficiency and accuracy of vision operations, at potentially lower costs, thanks to lower computational requirements.What are SSMs and how do they compare to transformers?SSMs are a neural network architecture class that processes sequential data differently from traditional transformers. While transformers use attention mechanisms to process all tokens in relation to each other, SSMs model sequence data as a continuous dynamic system.Mamba is a specific SSM implementation developed to address the limitations of earlier SSM models

DeepSeek introduced an upgrade to its artificial intelligence model.The new version of the Chinese startup’s V3 large language model was made available through AI development platform Hugging Face, Reuters reported Tuesday (March 25).The release marks DeepSeek’s latest effort to make a name for itself in an evolving AI sector, competing with the likes of OpenAI and Anthropic, the report said. The new model shows improvements over its predecessor in areas like reasoning and coding.DeepSeek rocked the tech world earlier this year when it unveiled a series of AI models that were said to perform at the same level as OpenAI’s ChatGPT but at a lower cost.Interviewed at a conference in China this week, Apple CEO Tim Cook reportedly described DeepSeek’s AI models as “excellent.”Asked about the risk of DeepSeek during Apple’s earnings call Jan. 30, Cook said: “In general, I think innovation that drives efficiency is a good thing. And that’s what you see in that model.”But while the debut of DeepSeek led observers to question the need for investment in AI infrastructure, it also prompted a greater focus on reasoning models, which require more spending on inference.AI investments by hyperscale companies such as Amazon, Meta and Microsoft will increase faster than earlier forecasts, with more of that money being spent on running AI systems once they have been trained, instead of on data centers and chips, Bloomberg Intelligence found.The data showed those companies spending $371 billion on data centers and computing resources in 2025 — 44% more than they spent last year — and $525 billion a year by 2032.Meanwhile, AI is being used to speed up the work of customer experience teams and make customers feel heard and better understood.“We’re finally moving beyond superficial ‘personalization,’” said Lisa O’Malley, senior director of industry products and solutions at Google Cloud, in a blog post last month. “AI-powered CX creates the feeling of being understood, of having needs anticipated and met with minimal effort.”For example, O’Malley said customers have begun saying “please” and “thank you” to AI agents, although the “most significant shift, however, is the evolution of the support system from a cost center to a revenue generator. The conversations I’m having with customers point to omnichannel engagement — across voice, web, mobile, email and apps — as directly driving ROI.”For all PYMNTS AI coverage, subscribe to the daily AI Newsletter

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More. Midjourney is best known as one of the leading AI image generators — with nearly 20 million users on its Discord channel, according to third-party trackers, and presumably more atop that on its website — but its ambitions are beginning to expand. Following the news in late summer 2024 that it was building its own computing and AI hardware, the company this week released a new research paper alongside machine learning experts at New York University (NYU) on training text-based large language models (LLMs) such as Meta’s open source Llama and Mistral’s eponymous source models to write more creatively. The collaboration, documented in a new research paper published on AI code community Hugging Face, introduces two new technieques — Diversified Direct Preference Optimization (DDPO) and Diversified Odds Ratio Preference Optimization (DORPO)— designed to expand the range of possible outputs while maintaining coherence and readability.For a company that is best known for its diffusion AI image generating models, Midjourney’s new approach to rethinking creativity in text-based LLMs shows that it is not limiting its ambitions to visuals, and that, a picture may not actually be worth a thousand words. Could a Midjourney-native LLM or fine-tuned version of an existing LLM be in the cards from the small, bootstrapped startup? I reached out to Midjourney founder David Holz but have yet to hear back.Regardless of a first-party Midjourney LLM offering, the implications of its new research go beyond academic exercises and could be used to help fuel a new wave of LLM training among enterprise AI teams, product developers, and content creators looking to improve AI-generated text.It also shows that despite recent interest and investment among AI model providers in new multimodal and reasoning language models, there’s still a lot of juice left to be squeezed, cognitively and performance-wise, from classic Transformer-based, text-focused LLMs.The problem: AI-generated writing collapses around homogenous outputsIn domains like fact-based QA or coding assistance, LLMs are expected to generate a single best response