It's telet multi-communication. We are committed to research and development of next-generation efficient AI models.
inquiryModern AI development faces serious structural challenges due to the oligopolization of massive models and the computational resources that support them. The situation where many talented researchers and developers are constrained by the latency and costs of black-boxed APIs, unable to access model internal behavior, represents nothing less than a loss of innovation opportunities. We view this as the loss of "computational sovereignty" caused by technical and economic barriers.
We clearly rebel against this centralized paradigm. Based on the unwavering belief that the true value of AI lies not in the enormity of parameter counts but in its "resource efficiency," we lead the next generation AI revolution. Our mission is to liberate AI execution environments from large data centers and make them ubiquitous across edge, mobile, and even personal devices. This is not merely a story of cost reduction, but a fundamental challenge to achieve true democratization of AI development and create new application ecosystems.
The core of our research and development lies in advanced technological clusters that maintain or improve model accuracy while dramatically reducing computational costs (FLOPs) and memory footprint by orders of magnitude. Specifically, beyond static structured pruning and non-uniform quantization (INT4/INT2), we pursue Dynamic Sparsification that dynamically optimizes computational paths during inference. Going beyond knowledge distillation from a single teacher model, we explore mechanisms for models to become smarter through mutual learning and self-distillation. And the design and implementation of linear-time attention mechanisms using kernel methods and low-rank approximations to break through Transformer's quadratic computational complexity. These are just a few examples of the challenges we tackle.
What these technological breakthroughs open up are application domains that were previously products of fantasy. Interactive systems where real-time performance is an absolute requirement, field agents that autonomously make situational judgments in offline environments, and above all, personal AI that operates while maintaining complete privacy in users' hands. This vast and unexplored territory is precisely the "future margin named computational efficiency" that we explore. Drawing that map first and architecting new standards for AI development. That is the core of the research and development we pursue.
Standard vector search is high-cost and high-latency. We build "hierarchical semantic clustering" that dynamically limits vector space according to queries. By combining coarse-grained cluster exploration with detailed search within those subspaces, we reduce computational complexity at the order level, enabling high-speed, high-precision search even with small-scale resources.
LLM's autoregressive token generation contains redundant computational paths. We research "inference path distillation" where expert decision-making logic is pre-trained into the model as computational graphs. By combining this approach with INT8/INT4 quantization, we avoid zero-based thinking during inference and maximize response performance.
Monolithic large models have reliability issues such as hallucination. We design lightweight LLM agent groups specialized for each task such as search, analysis, and verification. Knowledge is shared and coordinated on a distributed blackboard system, generating robust and reliable output through consensus formation.
Lightweight High-Speed Inference Engine
Collaborative Multi-LLM System
Contact WhatsApp #okuizumi keita
Contact via WhatsApp