multimodal AI

German AI company Aleph Alpha has released a technical report on Kolibri, its multimodal foundation model. The PDF, published on the company's website, is being discussed by technology readers, with many weighing how the German challenger's approach and capabilities compare to larger US-based AI labs.

Google has released EmbeddingGemma 2, a new embedding model the company says can run directly on devices and handle multimodal inputs. The launch is being discussed among developers and AI watchers as part of the push to bring capable AI models to phones and other consumer hardware without relying on cloud servers. Reaction online has focused on what the model means for privacy, latency and on-device applications.

A new paper introduces UniEvo-VL, a multimodal AI model trained through self-distillation, allowing it to improve its own performance without relying on large amounts of externally labeled data. The approach is being discussed among researchers as an example of growing interest in self-improving model training methods, and the paper is available on arXiv.

TwelveLabs has released Pegasus 1.6, a video understanding model aimed at physical AI applications such as robotics. The company says the model can analyze video input to help machines and robots interpret real-world visual environments, extending multimodal AI beyond screen-based tasks into embodied systems operating in physical spaces.

French AI company Mistral has released Large 4, a multimodal model with one trillion parameters, of which 49 billion are active at inference, jokingly nicknamed le Chonk. The company says it trained the model on 3,800 NVIDIA GPUs in its European data centres and reports an 82% score on a vulnerability reproduction-and-patching test.

AI startup Reka AI has introduced Rho-1, an omni-model designed to process text, images, and video while also controlling robots, all within a single model. The announcement is drawing attention for combining multimodal understanding with real-world robotic control, a capability usually handled by separate systems. Observers are watching closely to see how Rho-1 performs against established multimodal AI offerings.

Researchers and commentators are debating how image understanding should be defined now that multimodal AI systems can interpret pictures. The discussion draws on visual epistemology, shifting the focus from what images contain to how seeing, knowledge, belief and interpretation interact to produce meaning, and questioning whether machines genuinely understand images or merely simulate recognition.

Google has announced EmbeddingGemma 2, a model designed to bring multimodal semantic search to edge devices. The release lets developers run embedding-based search across text, images and other media locally, without sending data to the cloud. It signals Google's continued push to move AI capabilities onto phones and other resource-constrained hardware.

Dalian University of Technology has introduced VA-Bench, a benchmark for evaluating multimodal AI models on robotics tasks. Results show that even the top-performing models complete only about half of the tested tasks, highlighting a significant gap between current multimodal capabilities and practical robot control. The benchmark is drawing attention as a measure of how far AI still is from reliable real-world robotics.

BostonGene's Chief Medical Officer is scheduled to speak at the 2nd Annual Global Cancer Research & Innovation Symposium, where the focus will be on biologically grounded artificial intelligence and the use of multimodal data in oncology drug development. The company argues that combining AI with biological context can improve how cancer therapies are discovered and developed.