Skip to content
Main Site News Console

Primitive Rhythm Releases NeoHorse-1.5-Flash Preview as Agent Model RSI Enters Second Iteration

· 量子位
国内AI

On November 5, TokenRhythm officially released its agent-native model, NeoHorse-1.5-Flash Preview. Trained on Qwen3.8-Flash, the model features 180B total parameters with 6B active parameters, delivering focused enhancements to agent task execution and software engineering capabilities. It is currently available for limited closed beta testing.

Image 3

According to official comparative benchmark data, the new model achieved a 9.23-point increase in its overall score on Agents’ Last Exam and a 5.92-point increase on Natural Language to Repository compared to the base model. Notably, scores in physical science tasks improved by over 20 points, with significant performance gains also recorded in technical domains such as developer tooling & testing, web development, and security authentication & encryption.

This release continues the NeoHorse series’ exploration of Recursive Self-Improvement (RSI), marking what TokenRhythm terms the “second cycle” of this evolutionary mechanism. Following the launch of NeoHorse-1 in September, the team integrated the previous-generation model back into the model pool to continuously gather dynamic feedback during real-world task execution, feeding this accumulated experience back into subsequent training. The model training scale has also expanded from the earlier 4B and 9B tiers to the Flash tier.

This evolutionary path relies on the TokenRhythm multi-model platform and its open-source agent framework, OpenSquilla. The routing system provides demand signals for task capabilities, while a multi-model fusion mechanism orchestrates multiple frontier models to collaborate on the same task, producing results for cross-comparison. Discrepancies between NeoHorse and these outputs are leveraged to precisely diagnose capability bottlenecks, guide future training directions, and serve as learning signals for autonomous self-iteration.

The day following the model release, TokenRhythm updated OpenSquilla to further refine this execution foundation. The new version can dynamically match model tiers based on task intent—retaining the original tier when resuming a task, and progressively escalating tiers when retrying. It also supports independent updates to recommended configurations for TokenRhythm and OpenRouter, allowing users to preview configuration diffs and apply changes without upgrading the client. In addition, the latest recommended configurations have been calibrated against the November model leaderboards, ensuring that runtime model invocations remain aligned with current industry benchmarks.

Centered around the RSI training paradigm, the team introduced three core technologies: multi-model differential diagnosis, adversarial environment evolution, and swarm-intelligence online learning—addressing capability diagnosis, distribution collapse, and iteration efficiency, respectively. Among these, the environment evolution engine autonomously and continuously generates and filters new tasks, while swarm-intelligence online learning utilizes the delta between single-model outputs and multi-model consensus results to perform reinforcement training, efficiently converting comparative data gathered across the execution framework into an engine for model optimization.

Public demonstrations highlighted the new model’s performance leap across practical business tasks. In a spreadsheet web application development showcase, the base model produced formula errors in several calculated cells, whereas the new model not only flawlessly implemented formula inputs and calculation logic but also introduced richer interactive features. Official release materials also showcased use cases spanning chart generation, quantum physics computation, and robotic assembly parameter solving, thoroughly evaluating the model’s real-world engineering readiness across computation, validation, and end-to-end delivery.

Applications for the NeoHorse-1.5-Flash Preview closed beta are now open via TokenRhythm’s official WeChat account.