TokenRythm (Jiyuan Ludong, 极元律动), a Chinese AI startup founded by former Huawei Noah’s Ark Lab director Wang Yunhe, has released NeoHorse-1, its first Agent-Native model. The new release is available in 4B and 9B versions and is designed for the practical sequence of tasks that agents must handle, including tool use, reading environmental feedback, recognizing errors, revising a path, and completing work. QbitAI’s September 8 report describes the model as an effort to convert observations from multi-model agent execution into training material.
The release is notable because the company did not begin as a conventional foundation-model maker. TokenRythm has emphasized routing and orchestration systems that decide which models should handle parts of a task. NeoHorse-1 takes a step beyond that layer. It tries to use the traces generated while agents perform real tasks to improve a model’s own ability to work through those tasks.
That idea sits at the center of a changing agent market. As more models become available, companies need ways to select among them, control costs, recover from failures, and make sure an agent reaches an outcome rather than merely produces a plausible answer. The key question for NeoHorse-1 is whether that operational knowledge can produce reliable gains when moved into post-training.
A model shaped by execution traces
QbitAI said NeoHorse-1 combines signals from the company’s Routing Harness with public data. An agent run can produce far more than a prompt and an answer. It can record which model was selected, what tool calls it made, what the environment returned, whether an action failed, and how the system attempted a recovery.
Those records can be valuable, but they are not automatically useful training data. A long trace may contain irrelevant system messages, repetitive failed attempts, or a result that looks complete but did not satisfy the actual request. QbitAI reports that the startup evaluates traces against multiple dimensions, including whether the user goal was completed, instructions were followed, tools were used appropriately, results had supporting evidence, and errors were recovered from.
The model’s stated focus reflects the difference between an agent and a chatbot. A chatbot can provide an informative response without changing anything outside the conversation. An agent may need to inspect files, invoke tools, interpret results, change course, and leave a usable output. Training for that process requires attention to sequence, feedback, and verification.
TokenRythm’s approach resembles the broader interest in agentic post-training across the industry, but it is not a guarantee of autonomous improvement. The company’s public account describes one engineered loop in which human-designed systems collect, evaluate, select, and use execution evidence. It should not be confused with a demonstration that a model can recursively improve itself without human direction.
Routing becomes a source of training signals
The company calls one part of its method Routing-Guided Curriculum. In simple terms, routing signals can estimate the difficulty of a task and help order examples from less demanding to more complex ones. That may make training more efficient than treating every trajectory as equivalent.
QbitAI also reports that the project uses On-Policy Distillation. In this approach, a stronger model can provide guidance on the kinds of steps a smaller model actually takes, rather than only on a prewritten set of ideal answers. The aim is to help a model improve in situations it is likely to encounter during execution.
The 4B version is particularly important to the company’s cost argument. QbitAI reported that the model’s macro-average score rose from 58.94 to 64.87 after agentic post-training. That is a company and technical-report result, not an independent benchmark verdict, and it depends on the selected tasks and evaluation setup. Still, it illustrates what smaller models may be able to do if they are trained for a limited class of observable, verifiable workflows.
China’s AI market has increasingly focused on the economics of agents, not just the scale of base models. Earlier EastFrontier coverage of Alibaba’s Qwen model work for enterprise agents examined how Chinese companies have been adapting models for business use. NeoHorse-1 approaches the same economic question through orchestration: if a smaller model can reliably perform part of a workflow, the system may save calls to larger and more expensive models.
The challenge beyond a promising benchmark
NeoHorse-1 arrives in a competitive environment. Leading model developers are building their own agent tools, routing systems, and enterprise products. The challenge comes as Chinese open-source AI models have gained global download traction, expanding the number of systems an agent platform may need to evaluate and coordinate. Infrastructure providers also want to offer unified access to multiple models. A startup that sits between model suppliers and applications must show that its layer does more than make API access convenient.
TokenRythm argues that its advantage comes from cross-model evidence accumulated by its routing system. The claim is plausible in principle: a system that observes several models attempting similar tasks can learn about their strengths, weaknesses, and recovery paths. But the quality of that advantage will depend on the diversity of tasks, the integrity of the data, customer consent and privacy controls, and the ability to turn those records into improvements that persist outside a single benchmark.
The company’s model release therefore marks a useful test. It is not just another small model entering an already crowded market. It is an attempt to make the operational byproducts of agent systems into an asset for model training. Whether that becomes a durable advantage will depend on evidence from real deployments, not only on a published score.
For now, NeoHorse-1 shows where China’s AI startups are looking for differentiation. The race is no longer only about releasing larger models. It is also about building the systems that decide how models work together, capture lessons from their failures, and turn those lessons into more capable and less costly agents.