CaLICO-ROCK, a Real Operational Context Keystone for CaLICO LLMs

CaLICO is a family of European-developed large language models purpose-built for multilingual harmful language processing — spanning hate speech, disinformation, and extremist rhetoric across online platforms. Originally developed with support from an AI-BOOST award at European HPC centres, CaLICO demonstrated that a domain-specialised, sovereignly deployable model can match frontier commercial systems on harmful language tasks across 24 languages. What it currently lacks is the operational architecture to deploy that capability flexibly across diverse real-world use cases.

CaLICO-ROCK addresses this directly. Its central contribution is a replicable framework for composable, sovereignly deployable adapter ecosystems for sensitive-domain agentic pipelines — enabling role-specific analytical tasks such as disinformation detection, bias analysis, intelligence work, and media monitoring to be served from a single model base, without full retraining, within sovereign or air-gapped infrastructure. This is the missing layer between a capable foundation model and an operationally useful one, and no comparable open European model currently provides it.

Delivering this framework requires two enabling technical advances. First, continued pretraining on latest data to incorporate recent geopolitical developments and shifting rhetoric patterns, keeping CaLICO current at 3B–70B-parameter and MoE scale. Second, extended context windows and explicit reasoning capabilities, enabling native long-document analysis and auditable, knowledge-grounded outputs — the analytical complexity that real-world disinformation work demands. Together, these advances underpin a system that can operate as both a mission-critical platform for public sector security and a scalable commercial product across the information integrity market.