TokenRhythm unveils NeoHorse-1.5-Flash preview, boosting agent capabilities

TokenRhythm has made the NeoHorse-1.5-Flash Preview available in a limited internal beta. This agent-native model, which is built upon Qwen3.8-Flash, features a total of 180 billion parameters with 6 billion active parameters, specifically engineered to enhance software engineering and agent task execution. Performance evaluations indicate significant gains over the base model, including a 9.23 point increase on Agents’ Last Exam, a 5.92 point rise on Natural Language to Repository, and improvements exceeding 20 points in physical science tasks. The model also shows enhancements across developer tools, web development, and security certifications. This release is part of the NeoHorse series' ongoing investigation into recursive self-improvement (RSI), which TokenRhythm terms the second loop of this process. Following the September release of NeoHorse-1, the company reincorporated the older model into the training pool to gather execution feedback, expanding the training scale from the initial 4B and 9B tiers to the Flash tier. This methodology leverages TokenRhythm’s multi-model platform and the open-source agent product OpenSquilla, which uses routing signals to pinpoint capability gaps. In tandem with the model launch, OpenSquilla received an update allowing users to select model tiers based on task intent and to escalate tiers during retries, all without requiring a client upgrade.