Apodex-1.0 tiny open models (0.8B–4B) built for agent verification
Three very small open-weight language models have been released, each fine-tuned for the job of verifying AI agent outputs. The smallest is just 0.8B parameters — small enough to run for free on a laptop. They offer a cheap local replacement for expensive large-model API calls in agent pipelines.
In multi-step AI agent systems, a 'verification' step checks whether the agent's output is actually correct before passing it along. Today that step often uses large, costly models like GPT-4. Apodex-1.0 focuses entirely on doing that one job well, coming in three sizes — 0.8B, 2B, and 4B parameters — all small enough to run on consumer hardware at no API cost.
Performance was measured using AgentHarness, a benchmark specifically designed to evaluate agents. Because the weights are fully open, anyone can download, modify, and plug these models into their own pipelines. For teams running many agent calls, swapping the verification step to one of these local models could cut costs significantly.
Key points
- Three sizes (0.8B, 2B, 4B) — all small enough to run locally for free
- Specifically trained for the verification step in AI agent workflows
- Fully open-weight: download, modify, and integrate freely
- Benchmarked with AgentHarness, an agent-focused evaluation suite
- Replace expensive large-model API calls with a cheap local verifier
Quick term guide
- open-weight
- The model's internal numbers are publicly released, so anyone can download and run or modify it freely.
- parameters
- The internal numbers an AI model learns during training; more parameters generally mean a more capable model.
- placement
- The process by which a university helps graduates find jobs at companies, often measured as a percentage of students hired.
- AgentHarness
- A testing tool built specifically to measure how well AI agents perform on real tasks.
- benchmark
- A test used to compare speed, quality, or cost.
- local model
- An AI model you run directly on your own computer, with no internet connection or external service needed.
- agent workflow
- A set of steps an AI follows automatically to complete a series of tasks in order.
- valuation
- The amount investors think a company is worth.