TokenRhythm Releases NeoHorse-1.5-Flash Preview as Its Agent Model’s RSI Enters a Second Iteration
On November 5, TokenRhythm released NeoHorse-1.5-Flash Preview, an Agent-native model.
On November 5, TokenRhythm released NeoHorse-1.5-Flash Preview, an Agent-native model. Trained on Qwen3.8-Flash, the model has 180B total parameters and 6B active parameters, with a focus on improving Agent task execution and software engineering capabilities. Limited beta testing is now open.
According to the comparative evaluation results announced by the company, compared with the base model, the new model improved its overall score on Agents’ Last Exam by 9.23 points and its overall score on the Natural Language to Repository evaluation by 5.92 points. Performance on physical science tasks improved by more than 20 points, while areas including development tools and testing, web development, security authentication, and encryption also showed improvements.
This release continued the NeoHorse series’ exploration of recursive self-improvement (RSI), which TokenRhythm described as the mechanism’s “second cycle.” After releasing NeoHorse-1 in September, the company returned the previous-generation model to the model pool, collected execution feedback from real-world tasks, and used those experiences in subsequent training. The model training scale also expanded from 4B and 9B in the first round to the Flash tier.
This approach relies on TokenRhythm’s multi-model platform and the open-source Agent product OpenSquilla. The routing system provides signals about task capability requirements, while multi-model fusion enables multiple frontier models to handle the same task, producing comparable execution results. The differences between NeoHorse and these results are used to identify capability gaps, determine training directions, and convert them into learning signals.
On the day after the model’s release, TokenRhythm also updated OpenSquilla to improve this execution foundation. The new version selects a model tier based on task intent, retains the original tier when continuing a task, and progressively moves up a tier when retrying one. It also supports independently updating the recommended configurations for TokenRhythm and OpenRouter, allowing users to preview the differences and confirm their application without upgrading the client. The recommended configurations for this release have also been adjusted based on the November model leaderboard, enabling model selection on the execution side to update continuously in line with evaluation results.
For RSI training, the team proposed three techniques: multi-model differential diagnosis, adversarial environment evolution, and swarm-intelligence online learning, addressing capability diagnosis, distribution collapse, and iteration efficiency, respectively. The environment evolution engine continually generates and screens new tasks, while swarm-intelligence online learning uses the differences between model outputs and fusion results for training, applying the comparative information generated within the execution framework to subsequent optimization.
The public demonstration highlighted differences in specific tasks. In the spreadsheet webpage example, some calculation cells generated by the base model contained formula errors, while the new model demonstrated formula entry, calculation results, and additional interactive features. The release materials also listed cases involving chart generation, quantum physics calculations, and robot assembly parameter solving, assessing the models’ completion across calculation, checking, and delivery.
Applications for the NeoHorse-1.5-Flash Preview closed beta are now available through TokenRhythm’s official WeChat public account.
This article was provided by TokenRhythm and authorized for republication by QbitAI. All views belong to the original author.
