Inside Clark
The architecture behind the flagship
Clark is AICONSORTIUM's flagship multimodal intelligence system, built on a 70B backbone, scaled through Mixture-of-Experts routing, and deployed via the Federated Mesh. This is how it works.
Inside Clark
The architecture, in the open
A guided tour of Clark’s backbone, expert routing, and the federated mesh, and the design rationale behind the flagship.
Explore Clark
Overview
What Clark is
Clark is a multimodal large language model built from first principles for sovereign deployment. Its 70B dense backbone provides the reasoning foundation; a Mixture-of-Experts layer scales effective capacity toward 1 trillion parameters without proportional compute costs; and the Federated Mesh enables inference to be distributed across geographically sovereign compute nodes without compromising context fidelity.
Clark understands and generates text, code, structured data, and images, with native fluency across 100+ languages, including deep coverage of underserved and low-resource languages. It is the first large model designed from the ground up to be deployed on sovereign infrastructure.
70B
Dense backbone parameters
~1T
Effective MoE parameter count
100+
Languages supported
4
Modalities: text, code, data, image
Architecture
Four layers of intelligence
Clark's architecture stacks four distinct systems, each designed independently and integrated carefully.
70B Backbone
A transformer-based dense language model with 70 billion parameters, pre-trained on a curated multilingual corpus spanning major world languages and a deep set of low-resource languages. The backbone provides Clark's core reasoning, world knowledge, and language understanding capabilities.
MoE Routing
A sparse Mixture-of-Experts layer sits atop the backbone, routing each token to the two most relevant expert networks from a pool of specialised sub-networks. This scales effective model capacity to approximately 1 trillion parameters while keeping active compute per token fixed, enabling frontier capability without frontier compute costs.
Federated Inference
The Federated Mesh enables Clark's inference to be distributed across sovereign compute nodes, whether in government data centres, private enterprise infrastructure, or certified cloud environments, without any single node seeing the full context. Cross-node synthesis produces coherent outputs while respecting data jurisdiction boundaries.
Multimodal Fusion
Clark processes text, code, structured data, and images within a unified representation space, enabling cross-modal reasoning rather than simple modality switching. A dedicated vision encoder and a cross-modal attention layer allow Clark to reason about visual and textual content jointly, without needing separate prompting strategies per modality.
Design decisions
The choices that define Clark
Every architectural decision reflects a deliberate trade-off. Here are the three that most define what Clark is.
Why federated?
A centralised inference architecture requires that all data leave its jurisdiction to reach the model. For government institutions, healthcare providers, and legal systems, this is not merely inconvenient. It is constitutionally problematic. Federated inference means Clark can be deployed where data sovereignty laws require it, without architectural compromise.
Why MoE?
A fully dense model at 1 trillion parameters would require compute costs that limit access to organisations with the largest infrastructure budgets. MoE routing lets us achieve the capability of a 1T dense model while keeping per-request compute equivalent to a much smaller system, making frontier intelligence accessible at sovereign scale.
Why multilingual-first?
The world's linguistic diversity, with its thousands of languages, countless dialects, and multiple scripts, represents the hardest NLP challenge there is. Building Clark to native fluency across underserved and low-resource languages produces a model that is genuinely capable where every other frontier model degrades catastrophically. Multilingual-first is not market strategy. It is technical ambition.
Training Data
What we trained on, and why it matters
Clark's pre-training corpus was assembled with the same intentionality we apply to architecture design. We do not scrape the web indiscriminately and call it data. We curate: prioritising licensed content, government and institutional corpora, and high-quality multilingual text that reflects the linguistic reality of billions of people worldwide.
Low-resource language coverage was not achieved by translating English text. Clark's training includes substantial original-language content across more than 100 languages, from world majors to regional mother tongues, along with the code-switching patterns that characterise how multilingual speakers actually communicate.
We publish our data provenance methodology and maintain a public data card. Any institution deploying Clark on sovereign infrastructure can audit the training data composition for their deployment configuration.
Documentation
Read the full technical documentation
The Clark documentation covers architecture in full technical detail, including training methodology, evaluation benchmarks, known limitations, and deployment requirements.