Apodex 1.1 targets working capability, not just larger language modeling

The arXiv paper (2608.23283) defines working capability as continuous, verifiable progress toward a real-world goal, and the new model family is built around that definition. Instead of relying solely on reasoning and knowledge synthesis, the system combines persistent interaction with files, data sources and executable code with state maintenance, failure recovery and verifiable delivery.

Two scaling vectors

Development proceeds along two complementary axes. Environment Scaling expands the variety and fidelity of execution environments, files, search and code so the agent operates under conditions that resemble real professional work. Agentic Coordination Scaling trains agents to decompose long-horizon tasks, delegate work in parallel, integrate asynchronous results and replan when things go wrong. A shared execution harness and AgentOS preserve task state and data provenance across tools and agents, and training converts environment trajectories and coordination traces into reliable behavior.

Performance with a smaller model

According to the report, Apodex 1.1 reaches the top performance band across a range of domains — complex professional work, finance, scientific research, mathematics, coding and deep search — while using a model significantly smaller than many frontier systems. The mini version, at 35 billion parameters, retains strong working capability in a locally deployable form, opening the door for organizations that cannot or will not send data to the cloud.

Open workspace for research

The release includes three components available now: the full Apodex 1.1 through an online workspace, Apodex 1.1 mini with open weights for local execution, and FrontierAgent, an open-source research workbench that also incorporates an asynchronous agent team mechanism allowing human intervention and guidance at any stage. Code and weights are available in the repositories linked from the announcement.

Broader research context

The paper joins a wave of recent publications addressing long-horizon agents: Argus, LongHorizon-Harness, AstronOS, StructAgent, Agentao and others, all from 2026, signaling a clear direction in the research community — a shift from models that answer questions to systems that perform sustained, traceable and correctable work.