Apodex 1.1 aims at sustained work rather than benchmark scores

The company is launching a family of models that emphasizes long-term agent coordination and verifiable execution environments, and it claims front-running performance with a model markedly smaller than competitors.

The development focuses on two complementary dimensions: Environment Scaling and Agentic Coordination Scaling. The first dimension expands the variety and ability to verify actions in files, search and code; the second trains the model to decompose long-term tasks, delegate parallel work to multiple agents, merge asynchronous results and replan when required. A shared execution runtime and AgentOS preserve task state and data source across tools and agents, and the system turns environment pathways and coordination traces into reliable behavior.

According to the report, Apodex 1.1 reaches the leading performance band on complex professional tasks, finance, scientific research, mathematics, code and deep search, while using a model significantly smaller than many front-running systems. The mini version, with 35 billion parameters (35B), retains strong capability in a locally runnable form. The company defines the target as a “Heavy-Duty Solver” for ambitious, long-running tasks.

The launch includes three publicly available components: the full Apodex 1.1 via an online workspace, Apodex 1.1 Mini with open weights for local execution, and FrontierAgent, an open-source research environment that enables running the asynchronous agent team locally. The paper is available on arXiv under identifier 2608.23283.

The source does not publish numeric performance metrics, specific benchmark names or concrete comparison baselines; the claim of a “leading band” remains externally unverified. The paper is a pre-print that has not undergone peer review, and as of publication no other model cites it. The distinction between “open weights” and “open source” is relevant here: only the mini model and environment are released for local use, while the full model remains behind an online service.