Meta researchers taught an 8B AI model to match Claude Opus 4.5 — without the frontier price tag
Meta AI and University of Illinois researchers introduced EvoHarness‑RL, a framework that teaches an 8‑billion‑parameter model to manage its own belief, progress, and experience workspace, enabling it to perform on par with Anthropic’s Claude Opus 4.5 without the associated cost. EvoHarness‑RL adds a dynamic memory layer that can read, update, compress, and replace information during long‑horizon tasks, replacing rigid, manually‑coded harness logic. The approach lets engineers upgrade models without re‑engineering prompt and memory scaffolding, dramatically reducing engineering overhead for enterprise‑scale agents.
⚡ Key Takeaways
- 8B‑parameter model achieves performance comparable to Claude Opus 4.5.
- EvoHarness‑RL introduces a unified workspace that dynamically manages belief, progress, and experience.
- Dynamic memory avoids the “append‑only” degradation seen in static memory systems, improving long‑task reasoning.
- Integration replaces manual harness scripts with a learned abstraction layer that adapts to each model’s needs.
- Requires replacing existing static harness logic with the EvoHarness‑RL framework; otherwise, upgrades will still need manual retuning.
Engineers can deploy high‑performing, cost‑effective agents for complex, multi‑step enterprise workflows while cutting the time spent on custom harness engineering and memory management.
✅ Practical Steps
- Replace your agent’s static harness with the EvoHarness‑RL framework as described in the paper.
Want the full story? Read the original article.
Read on VentureBeat AI ↗