Ontelic creates next-generation language models — from lightweight assistants to large-scale deep analysis systems. We don't follow trends. We build the foundation.
Ontelic creates next-generation language models — from lightweight assistants to large-scale deep analysis systems. We don't follow trends. We build the foundation.
Ontelic isn't just another startup wrapping someone else's API. It's a fully-fledged research lab where every model is designed, trained, and optimized in-house.
Ontelic was founded with a single goal: to prove that high-quality artificial intelligence can be developed outside the reach of giant corporations, while surpassing them in efficiency, speed, and depth of understanding. We reject cookie-cutter solutions and build architectures that truly solve user problems.
In mid-summer 2026, the Ontelic mobile app will be released with an integrated chatbot — a personal AI assistant based on our models. Users will have access to Orynt Lemma and Axiom directly from their phones, with full context and synchronization.
Our philosophy is simple: the model must be predictable, safe, and useful. Each Orynt line undergoes multi-stage testing for sustainability, ethics, and accuracy. We don't chase hype — we chase results.
We don't create models from scratch or maintain our own servers. All our models are fine-tuned versions of popular neural networks: we take proven open-source architectures and transform them into specialized tools for specific tasks. We customize components, refine training on our own datasets, optimize inference, and implement our own security mechanisms. This is real R&D — not in a lab setting, but designed to solve specific user problems.
Each model in the Orynt line is designed for a specific use case. From instant answers to deep, multi-level analysis, choose the right tool for your needs.
Orynt Lemma, Orynt Axiom, and the Orynt Credo will be available to users simultaneously—from lightweight assistants to large-scale deep analysis systems. We are completing final testing and preparing the infrastructure for the full launch.
A lightweight and lightning-fast model based on a popular core for everyday tasks. Built for scenarios where every millisecond counts — from quick clarifications to short dialogues.
Orynt Lemma — is the industry's answer to truly fast AI. In a world where users expect instant response, Lemma delivers answers in fractions of a second without sacrificing quality and relevance. The model is optimized to run on a wide range of devices and infrastructures, making it the ideal choice for integration into mobile apps, chatbots, and embedded systems.
Architecture Lemma is built on principles of efficient attention and aggressive quantization without loss of semantic precision. We spent months fine-tuning every layer so the model understands user intent from the first tokens and formulates answers without excessive verbosity. Lemma doesn't write essays where a single sentence suffices — and that is its strength.
Despite its compactness, Lemma maintains a high level of safety and ethics inherited from the senior models in the lineup. It is resilient to jailbreak attempts, does not generate harmful content, and adheres to responsible AI principles. Lemma — is proof that small size does not mean a compromise in quality.
A balanced model based on a popular core with deep contextual understanding. Elaborate answers, resilience in long dialogues, and the ability to maintain the thread of conversation across thousands of tokens.
Orynt Axiom occupies the golden mean in our lineup. This is the model for those who value balance between speed and depth. Axiom is capable of conducting lengthy dialogues, maintaining context and nuances throughout the entire conversation. It understands subtext, recognizes irony, and can adapt its communication style to the interlocutor.
Architecture Axiom includes enhanced mechanisms of long-term memory and hierarchical attention, allowing it to efficiently process lengthy documents, technical specifications, and complex instructions. The model is undergoing final testing for resilience to hallucinations and the ability to admit ignorance — qualities we consider critically important for user trust.
Axiom will become a universal tool for business, education, and creativity. Its ability for critical analysis, summarization of complex texts, and generation of structured answers makes it ideal for integration into corporate systems, educational platforms, and research tools. It is expected that Axiom will set a new standard for balanced language models in its class.
A model based on a popular core with 671 billion parameters. Maximum depth of analysis, highest accuracy, and unique specialization in cybersecurity as White Hat AI.
Orynt Credo — is the pinnacle of our technological pyramid. With 671 billion parameters, this model is a monument to engineering thought and years of research. Credo is built for tasks where every nuance matters: legal analysis, scientific research, complex mathematical proofs, code auditing, and deep philosophical discourse.
Unique feature Credo — its specialization as White Hat AI. The model is trained on a massive corpus of cybersecurity, ethical hacking, and information protection data. It is capable of analyzing vulnerabilities, proposing defense strategies, conducting security audits of systems, and training specialists in the fundamentals of cybersecurity. At the same time Credo has strict ethical constraints: it will never use its knowledge to create malicious software or attack systems.
Architecture Credo includes advanced techniques of Mixture of Experts (MoE), dynamic attention, and multilevel answer verification. Every inference of the model undergoes internal consistency and factual accuracy checks. Credo is capable of engaging in multi-hour conversations on abstract topics while maintaining logical coherence and not contradicting itself. This is not just a language model — it is a deep thinking system.
Training Credo is conducted on a cluster of x4 H200, which allows us to realize the full potential of its massive architecture. We invest in infrastructure because we believe: the future of AI belongs to models that don't just answer questions but truly understand the essence of what is happening.
Leili.1 — is not a language model. It is a specialized system for visual creativity, fine-tuned to the limit. Image editing, graphic generation, working with maximum quality settings — Leili.1 is built for those who don't accept good enough".
Spot correction, style change, retouching and restoration at the level of professional tools. Leili.1 understands image context and preserves its integrity during any manipulation.
Operation at peak quality parameters without performance degradation. Every pixel is calculated with maximum precision to achieve photorealistic results.
The model analyzes artistic composition, color harmonies, perspective, and lighting. The results are not just beautiful — they are visually plausible and aesthetically refined.
From concept art for games to medical visualization — Leili.1 adapts its approach to a specific area of application, taking into account specific requirements for accuracy and style.
Leili.1 has undergone an intensive fine-tuning process on specialized high-resolution datasets. Unlike general-purpose multimodal models, Leili.1 is honed exclusively for visual tasks, allowing it to achieve results unattainable for systems of general purpose. It understands complex prompts including technical requirements for lighting, materials, cameras, and post-processing. For designers, artists, architects, and game developers Leili.1 becomes not a replacement for creativity but its amplifier — a tool that takes over the routine and frees the creator for true art.
Nat Drexov — founder, chief architect, and sole developer Ontelic. Over five years of self-study in cybersecurity, programming, and neural network architectures, he went from a forum reader to a creator of models with hundreds of billions of parameters.
Without formal education in the field computer science, Nat learned to program through books and forums, studied system vulnerabilities, analyzed exploits, and gradually came to an understanding: the future belongs to those who can create AI that doesn't just imitate intelligence but truly helps people solve complex tasks.
It was this cybersecurity background that determined the unique specialization Orynt Credo as White Hat AI. Nat believes that the most powerful models must be on the side of defense, not attack. This ethical stance permeates the entire philosophy of Ontelic: technology should amplify humans, not replace them.
Today Nat leads all directions of the company — from research into model architectures to infrastructure solutions. His approach to development is characterized by an unwavering commitment to minimalism: remove everything superfluous, leave only what is essential. This philosophy is reflected in both product design Ontelic, and their technical implementation.
Every decision in model architecture passes through the prism of this philosophy. There is no room for random layers, redundant parameters, or hacks for the sake of development speed. If a component does not bring direct benefit to the user — it is removed. This approach requires more time for design, but the result is systems that are predictable, safe, and understandable in operation.
We create AI that thinks deeper. That doesn't imitate — understands. That passes the test of predictability, safety, accuracy. But if the test is passed with a perfect score — and the response is only silence — does that mean the system is perfect or that a perfect system is not needed?Nat Drexov CEO Ontelic
Models Ontelic run on advanced clusters NVIDIA H200 — an architecture that sets the standard for modern computing for artificial intelligence.
Clusters based on NVIDIA H200 provide the necessary computing power for training and inference of models at scale Credo. Tensor cores and transformer engines accelerate every operation.
Infrastructure is built using NVLink and high-bandwidth interconnects, allowing distribution of 671B parameters Credo across multiple GPU without performance loss.
Distributed computing power without dedicated servers. Automatic scaling under load. Fault tolerance at every level. This is not rented capacity — this is a technological foundation built specifically for the tasks of artificial intelligence.
© Ontelic. We have no own serverss. All computations are hosted on the platform Modal.com. This gives us full control over data security, response latency, and equipment configuration H200 allow implementation of advanced distributed training techniques such as pipeline parallelism and tensor parallelism, at a level unavailable to most independent laboratories. Investments in hardware are investments in the speed of model thinking.
Important: when first launching models after downtime, response delays may occur due to cold-start — the time required to deploy and initialize computing resources. We are working to minimize these delays.