On November 5, TokenRhythm released the Agent-native model NeoHorse-1.5-Flash Preview. The model was trained on Qwen3.8-Flash, with 180B total parameters and 6B activated parameters, focusing on improving Agent task execution and software engineering capabilities, and is currently open for limited beta testing.
According to the company's published comparative evaluations, compared with the base model, the new model's overall score on Agents' Last Exam improved by 9.23 points, and its overall score on the Natural Language to Repository evaluation improved by 5.92 points. Among these, physical science tasks improved by more than 20 points, with improvements also in development tools and testing, Web development, security authentication and encryption, and other areas.
This release continues the NeoHorse series' exploration of recursive self-improvement (RSI), which TokenRhythm calls the "second lap" of this mechanism. After releasing NeoHorse-1 in September, the company put the previous-generation model back into the model pool, collected execution feedback in real tasks, and then applied this experience to subsequent training; the model training scale also expanded from 4B and 9B in the first round to the Flash tier.
This path relies on the TokenRhythm multi-model platform and the open-source Agent product OpenSquilla. The routing system provides signals of task capability requirements, while multi-model fusion allows multiple frontier models to process the same task, forming comparable execution results. Differences between NeoHorse and these results are used to identify capability gaps, determine training directions, and convert them into learning signals.
The day after the model release, TokenRhythm also updated OpenSquilla, improving this execution foundation. The new version selects model tiers based on task intent, keeps the original tier when continuing a task, and upgrades tier by tier when redoing tasks; it also supports independently updating the TokenRhythm and OpenRouter recommended solutions, allowing users to preview differences and confirm application without upgrading the client. This edition's recommended configurations were also adjusted according to the November model rankings, so that model selection on the execution side can be continuously updated following evaluation results.
For RSI training, the team proposed three techniques: multi-model differential diagnosis, adversarial environment evolution, and swarm intelligence online learning, corresponding respectively to capability diagnosis, distribution collapse, and iteration efficiency. Among these, the environment evolution engine continuously generates and filters new tasks, while swarm intelligence online learning trains on the differences between model outputs and fusion results, using the comparative information generated in the execution framework for subsequent optimization.
Public demonstrations presented differences in specific tasks. In a table web page case, some computed cells generated by the base model produced formula errors, while the new model displayed formula inputs, calculation results, and more interactive features. The release materials also listed cases such as chart generation, quantum physics computation, and robotic assembly parameter solving, examining the model's completion of computation, verification, and delivery stages.
NeoHorse-1.5-Flash Preview is now available for beta testing via application through TokenRhythm's official WeChat account.
This article was provided by TokenRhythm and reposted by QbitAI with authorization; the views belong to the original author.