We train, evaluate, develop and deploy AI systems. All our services are supported by Seeweb cloud provider. We ♥ Italy and open source.
Latest work
Discover our newest models and the tools to build what comes next. Open research, made to be explored.
Nev chooses. Nesso3 calls and answers. Nevso combines structured tool selection with small-model generation for local English and Italian agents. Your application validates calls and controls execution, with offline inference after setup on Linux and macOS.
Turn English and Italian requests into structured choices. Built on Nesso2, Nev selects among your supplied options for routing, function selection, and classification, returning typed decisions and option probabilities—not generated text or tool arguments.
Our latest bilingual Italian / English model for function calling and structured output. Built on Zagreus, Nesso3 learns a compact tool contract across thousands of tool schemas, with a focus on Italian tool use.
Research-informed fine-tuning with practical defaults. Memory-efficient training, native support for agentic data, and one configuration that scales from a single GPU to multi-node clusters.
Our research direction
Specialization is a research direction, not simply a way to reduce inference costs. We explore forms of intelligence that current frontier training objectives may not adequately prioritize.
Read our manifestoPurpose-built, efficient models for specific capabilities, languages, cultures, and domains. Specialization matters more than parameter count.
Model-to-model communication, orchestration, shared context, and multi-agent learning to bring complementary capabilities together.
Datasets and benchmarks that look beyond reasoning and coding to linguistic quality, cultural understanding, creativity, and cooperation.
Models, training recipes, evaluations, and tools that let the community reproduce results, experiment independently, and build on shared work.
We believe the future of AI lies in specialized models, each developing distinct capabilities and contributing to a broader intelligence through cooperation.
Research in focus
Meet the models and experiments behind our community. Built to be studied, shared, and taken further.
Bilingual English / Italian models designed to bring capable AI closer to the edge. Built on Zagreus, the Nesso family offers conversational instruction following, function calling, and structured output in a compact footprint.
Four bilingual foundation models, trained from scratch on around one trillion open-source tokens. Pairing English with Italian, Spanish, Portuguese, or French, Zagreus provides the foundation for Nesso—with open weights, configurations, and evaluations.
A 7-billion-parameter open model fine-tuned for fluent Italian across a wide range of domains. Maestrale brings together language generation, tool integration, structured JSON output, and reasoning capabilities.
An open research initiative exploring political bias in language models. Evaluation suites, datasets, and the Propaganda-dx and Propaganda-sx models help researchers study how political perspectives are represented and shaped.
Transparent evaluation for the Italian AI community.
General language evaluation
Compare open Italian language models on reproducible benchmarks, including MMLU IT, ARC IT, and HellaSwag IT.
Explore the leaderboard (opens in a new tab)
Domain-specific evaluation
Go deeper with specialized Italian benchmarks in logic, mathematics, law, foreign languages, and culture.
Explore the leaderboard (opens in a new tab)Built in the open
Models, datasets, training tools, and benchmarks. Explore the research we share with the ML and AI community.
Latest releaseBilingual decision model
An English / Italian decision-model family built on Nesso2 for routing, function selection, and classification over supplied options, with typed decisions and option probabilities.
Latest releaseSmall language model
Our latest compact bilingual Italian / English model for function calling and structured output, built for practical agentic workflows.
Training framework
A memory-efficient fine-tuning framework with native support for agentic data, scaling from a single GPU to multi-node training.
Code & research
Full technical report for the Zagreus & Nesso SLM families — from data pipeline and pre-training to post-training and evaluation — releasing 7 open bilingual Romance language models with fully reproducible configs.
Code & research
Propaganda is an open-source initiative dedicated to assessing and generating political bias in Large Language Models (LLMs).
Code & research
We contribute to add mmlu and arc_c evaluation benchmark for Italian based tasks. Our bechmarks are now used by many LLMs groups for evaluating Italian based LLMs
Models & datasets
One of the first completely open source Italian model in three versions: continual pre-trained, SFT and DPO
Models & datasets
The entire corpus of the current Italian laws as an open source dataset
Models & datasets
Translated version of the ultrafeedback dataset for SFT fine tuning
Models & datasets
Translated version eng to ita of the popular ultrafeedback binarized dataset for DPO fine tuning.
Models & datasets
The biggest dataset of real conversations in the Italian language from usenet
Code & research
We tranlated the mmlu-pro dataset to Italian and contributed to integrate into the lm-eval command for evaluating Italian LLM on difficult tasks
Models & datasets
A series of open source LLMs a fine tuned versions of mistral 7 Billion for producing Italian
Models & datasets
A multimodal and text dataset for evaluating LLMs on Italian tasks
Code & research
A fork of lm-evaluation-harness with Italian specific tasks: mmlu arc-c hellaswag pinocchio-law, pinocchio-logic, pinocchio-math and reasoning like mmlu-pro-ita