TokenRhythm 发布 NeoHorse-1，含 4B 和 9B 两个版本，基于 Qwen3.5 后训练，训练数据来自 OpenSquilla Harness 中智能体的结构化执行轨迹，涵盖路由决策、工具调用、失败与恢复步骤及结果。该模型通过 Data-RSI + Model-RSI 双循环实现自我改进，将执行经验反馈到训练中，并将通过 TokenRhythm API 提供。
Most agent systems throw away their most suitable training data after every run.
A task finishes, and the useful part disappears with it: which model got routed where, which tool was called, what came back, where the agent failed, and how it recovered.
And now, TokenRhythm released NeoHorse-1, a 4B/9B model family post-trained on agent execution traces and outcomes.
So it is trained on what an agent actually did, including its tool calls, mistakes, and recoveries, then that experience is fed back into training so the next model can perform better.
NeoHorse-1 is an early engineering validation of RSI through two connected loops: Data-RSI + Model-RSI.
Both the NeoHorse-1 checkpoints start from Qwen3.5.
The training loop around them uses structured execution trajectories generated inside
TokenRhythm's OpenSquilla Harness, capturing signals such as routing decisions, tool calls, failures, recovery steps, and outcomes. The updated model goes back into OpenSquilla to generate the next round of trajectories.
This is a really solid direction because, AI doesn’t automatically improve just because it uses tools. The key is whether execution experience can be captured, evaluated, and turned into useful training data.
Today, we’re introducing NeoHorse-1—our first in-house model, available in 4B and 9B variants. This is not a pivot away from what we have always believed: models will not converge into one winner. Agents need to organize models. NeoHorse is the next step in that thesis. It is trained not only on questions and answers, but on structured execution trajectories produced inside our Harness: when to search, when to use a tool, what a tool returned, what failed, how the system recovered, and what action should come next. The valuable signal is not only “how to succeed.” It is also where agents fail—and how they get back on track. This lets us close a foundational RSI loop inside OpenSquilla: Run tasks → record trajectories → filter learning signals → train the model → run a better model → collect better trajectories. To be precise: we have closed the basic Data-RSI + Model-RSI loop. We have not yet solved the harder problem of models autonomously designing their own improvement strategy. NeoHorse is still young. It is not here to replace every other model. It is an Agent-Native model: built to work inside a Harness, and to take its place in a broader network of specialized models. The goal is not one model that does everything. It is a system that gets better at getting work done. NeoHorse-1 will be available through the TokenRhythm API. We’re excited to learn from what it becomes with the community. #AIAgents #AgentEngineering在 X 查看被引用的帖子
来源：Rohan Paul · x.com