Apodex Releases Smol Models for Agent Workflows

Apodex has released Smol models (0.8B, 2B, 4B) as open weights, designed to perform specific sub-tasks within agent loops. These models specialize in Cross-examination and Execute & Verify, reducing the inefficiency of using large models at every step.

Key features include:

• Agent-specific training: Optimized to treat external text as 'claims' to be verified rather than simple facts, and to check the structural accuracy of tool calls.
• AgentHarness release: An evaluation framework is also open-sourced alongside the models, allowing local testing of long-horizon agent workflows spanning 50+ steps.
• Performance metrics: The flagship model, Apodex-1.0-H, demonstrated strong agent-based reasoning capability with high scores on benchmarks such as DeepSearchQA (94.4) and BrowseComp (90.3).