Prometheus / Prometheus 2
model Your tags
Your notes
The LK Lab's (Minjoon Seo) open evaluator LMs. Prometheus (2023, Llama-2-Chat-13B fine-tune on the Feedback Collection) was the first open judge to approach GPT-4-level fine-grained evaluation; Prometheus 2 (2024, built on Mistral-7B with a Mixtral-8x7B variant via weight merging) unified direct assessment and pairwise ranking and remains the most-used open evaluator model (~26K monthly downloads of the 7B in mid-2026). Anchors the prometheus-eval org's live 2026 evaluator-agent push (K-BrowseComp Korean deep-research benchmark, peer-review benches).
Model Details
License Apache 2.0
Base model mistral-7b