TokenRhythm recently released the agent-native model NeoHorse-1.5-Flash Preview. Trained on Qwen3.8-Flash, it has 180B total parameters and 6B active parameters, with a focus on improving agent task execution and software engineering capabilities. It is currently available in a limited internal beta.
Comparative benchmarks show that, compared with the base model, the new model improves its overall score by 9.23 points on Agents’ Last Exam, by 5.92 points on Natural Language to Repository, and by more than 20 points on physical science tasks. It also improves in areas such as developer tools and testing, web development, and security certification and encryption.
This release continues the NeoHorse series’ exploration of recursive self-improvement (RSI), which TokenRhythm calls the “second loop” of this mechanism. After releasing NeoHorse-1 in September, the company put the previous-generation model back into the model pool to collect execution feedback for subsequent training, expanding the training scale from the first round’s 4B and 9B to the Flash tier.
This approach relies on TokenRhythm’s multi-model platform and the open-source agent product OpenSquilla. It uses routing signals and differences in multi-model fusion results to identify capability gaps and determine training directions. The day after the model release, OpenSquilla was updated in sync, adding support for selecting model tiers based on task intent, stepping up tiers on retries, and independently updating recommendation plans. Users can preview the application without upgrading the client.
