DeepMind alumni’s tiny 27B model crushes GPT-5.5 at scientific research
inherent's faraday agent, running on a 27b parameter model, outperformed anthropic and openai at replicating scientific papers, proving efficiency beats size in complex agents.
inherent emerged from stealth with a $50m seed round and benchmarks showing its faraday agent can replicate published papers without prior answers, beating anthropic claude opus 4.8 and openai gpt-5.5. the system runs on qwen 3.6, a 27b parameter model, leveraging reinforcement learning to acquire research taste and experimental instincts rather than just scaling parameters. this challenges the industry consensus that bigger is better, suggesting that specialized training methods enable smaller models to outperform giants on specific, high-complexity workflows. inherent focuses specifically on scientific discovery, using tools like gpt-5.5 codex for coding to avoid reinventing the wheel and positioning itself as an autonomous research teammate.