Skills
NVIDIA Partners With Nous Research on Hermes: Qwen 3.6-27B Outperforms Previous 397B Models on Agentic Coding — Optimized for Always-On Local Inference on DGX Spark's 128GB Unified Memory
NVIDIA's RTX AI Garage announced Hermes Agent optimization for DGX Spark (128GB unified memory, 1 petaflop AI) and RTX PCs on May 13. The key technical claim: Qwen 3.6 27B and 35B parameter models now outperform their previous-generation 120B and 400B counterparts on agentic coding benchmarks, making high-quality local agents practical on consumer and prosumer hardware. Hermes is provider- and model-agnostic, designed for always-on 24/7 local use. Combined with Hermes's self-evolving skill architecture, this positions local agents as a serious alternative to cloud APIs for developers who want persistent, private, and rate-limit-free agent access.
Source
↳ Follow the thread