27B AI Scientist Faraday Released
Inherent Labs has released Faraday, a 27B-parameter 'AI scientist' trained via long-horizon reinforcement learning specifically to reproduce research papers. Official evaluations show it outperforming Claude Opus 4.8 and GPT-5.5 on held-out paper reproduction tasks. If independently verified, the result would suggest models far smaller than flagships could catch up on scientific research tasks, and paper reproduction might become a new training paradigm.
Confirmed
- Inherent Labs launched Faraday between 08-14 and 08-15; at 27B parameters it is positioned as an 'AI scientist' and is the team's first public release of this kind
- The accompanying research proposes the Replica method, which turns paper reproduction into a scalable reinforcement learning task space, and Faraday was trained with long-horizon RL accordingly
- Faraday operates as a coding agent, which the team says adds 'scientific intuition' to coding agents
- Developer-run evaluations claim Faraday beats Claude Opus 4.8 and GPT-5.5 on held-out research reproduction tasks
Not yet confirmed
- All comparison results come from the developers' own evaluations and secondhand accounts (@Scobleizer also phrased it as 'reportedly'), with no independent third-party verification so far
- The post materials do not disclose benchmark details, training data composition, or whether the model is open-weight or API-only
Why it matters
- If the results hold, long-horizon RL with paper reproduction as the task space could let a 27B-scale model surpass frontier closed-source flagships on research tasks; @omarsar0 already views it as a new training paradigm
- Paper reproduction is a well-defined, verifiable part of scientific research; turning it into tasks makes rewards automatically judgeable, potentially supplying scalable training data and reward sources for the 'AI scientist' direction
2026-08-14 ~ 2026-08-15 · 6 related posts
Primary sources
- Faraday: 27B AI Scientist Outperforms Claude Opus 4.8 on Paper Replication — Tom_Westgarth15 · 2026-08-14
- Inherent Releases First Public AI Scientist: Faraday — Scobleizer · 2026-08-15
- [source] 27B agent Faraday beats Claude Opus 4.8 and GPT-5.5 on research replication via new Replica method — omarsar0 · 2026-08-15
- Inherent Releases Faraday: AI Scientist Outperforms GPT-5.5 in Replicating Papers — ricklamers · 2026-08-15
2 near-duplicate retellings: matthen2 · ChengleiSi