According to Leiphone, TokenRhythm released the agent-native model NeoHorse-1.5-Flash Preview on 11 5, trained on Qwen3.8-Flash. With 180B total parameters and 6B active parameters, it focuses on enhancing agent task execution and software engineering capabilities, and is currently in limited beta testing. The company's published comparative evaluations show that, compared to the base model, the overall score improved by 9.23 points on Agents’ Last Exam and by 5.92 points on Natural Language to Repository. The company states this is the "second loop" of recursive self-improvement for the NeoHorse series: placing the previous generation model back into the model pool to collect execution feedback from real tasks for further training.
Why It Matters Iteratively training agent models using execution feedback from its own multi-model platform, the company claims improved scores on engineering tasks.
Credibility: Fairly Credible · Source: Leiphone · Event Time: 11-05 (Beijing Time)
Reasoning
- A domestic startup applies recursive self-improvement to agent model training
Risks
- Evaluations are self-reported by the company; limited beta testing, no public API pricing yet