Nous Research releases Hermes 4.3, trained on Psyche network

Nous Research released Hermes 4.3, a 36-billion-parameter model built on ByteDance’s Seed-OSS-36B base with a 512,000-token context window, as the first flagship Hermes release trained entirely on Psyche, the company’s peer-to-peer distributed training network, rather than on a centralised GPU cluster.

Psyche coordinates training across geographically dispersed nodes using Nous’s DisTrO optimiser to handle gradient communication over a custom mesh network, with Solana used for consensus on training state — an arrangement Nous described as “local intelligence, globally trained.” To test the approach, Nous said it trained comparable models both centrally and on Psyche, and that the Psyche-trained version outperformed its centrally-trained counterpart on downstream tasks. The company also reported state-of-the-art results on its own RefusalBench, a benchmark it uses to measure model helpfulness and appropriate refusal behaviour rather than standard maths or coding performance.

Distributed training across commodity or geographically separated hardware has been a longstanding goal for open-source AI groups seeking to compete without access to hyperscale clusters; Hermes 4.3 was presented as evidence the approach could now produce a flagship-quality open-weight model rather than remaining a research demonstration.