The company reports a 61.8% result on Terminal-Bench 2.1 with thinking mode turned off.
For the same model, Cline lists scores of 53.9%, 51.7%, and 44.9% for three unnamed open-source harnesses. It says two of those runs used high reasoning or effort settings. The article describes Cline features including an instruction to plan before implementation, a stall detector that records structured failure events, and task-specific verification checks.
Cline also says it uses an internal Nemotron-based agent in Slack for task triage and investigation of user-reported bugs; the agent can create sandboxes and containers and call more advanced models. The post does not identify the comparison harnesses, provide full evaluation artifacts, isolate the contribution of any Cline feature, or measure token or compute savings. It also does not establish availability, pricing, or free access for Nemotron or other models.