NeoHorse-1 Operational

NeoHorse-1 is an open-source Agent model released by TokenRhythm in collaboration with Wuwen Xiongtu, Tsinghua University, Peking University, The Chinese University of Hong Kong, Alibaba, and others. It comes in 4B and 9B versions. The model uses Agent-Native post-training, with real execution trajectories from Routing Harness as training data. Routing signals are reused to determine the order of the training curriculum, creating a closed loop of evaluation feedback → data allocation → model updates. The model supports tool calling, task planning, deep search, and code generation. The 9B version achieves an average score of 69.04, exceeding its base model by 3.44 points. It natively supports a 260,000-token context and can be deployed locally.

Tool calling: Automatically selects and calls external tools based on task requirements to complete multi-step operations.
Task planning: Breaks complex tasks into executable subtasks and performs long-horizon planning and scheduling.
Deep search: Actively retrieves and integrates information over multiple rounds of interaction to support an evidence chain for conclusions.
Code generation: Generates usable code and debugs and fixes it based on environment feedback until the task is complete.

Follow WeChat and reply “开源” to join the AI open-source project discussion group
Harness trajectory-driven Agent-Native training: The training corpus centers on structured trajectories generated through real Routing Harness executions, covering requests, routing, tool calls, environment feedback, and error recovery. Compared with traditional question-and-answer data, it adds three dimensions: capability requirements, execution decisions, and environment results. Both successful and failed trajectories are retained. After structural checks and six-dimensional quality evaluation, selected trajectories are added to the training set so that the training objective directly targets task completion.
Capability-guided data allocation: The router divides tasks into four levels, C0–C3, according to capability requirements. The signals are reused offline as training criteria: routing records are divided into three stages, “prediction → action → result.” The prediction stage is used to arrange the curriculum, while the result stage identifies capability gaps. The next round’s data allocation is then adjusted accordingly, forming a closed loop of evaluation feedback → data allocation → model updates.
Single-round validation of recursive self-improvement (RSI): This consists of Data-RSI and Model-RSI. The former refers to the model continuously executing tasks to generate new trajectories, which are filtered and naturally accumulated as material for subsequent training. The latter refers to updating the model based on evaluation gaps and returning it to the Harness model pool. Currently, only a single closed loop has been validated; the signal and reward designs are still defined by humans.

Install the download tool: Run pip install -U modelscope to install the ModelScope command-line tool.
Download the model: Run modelscope download --model TokenRhythm/NeoHorse-1-9B --local_dir ./NeoHorse-1-9B to download the weights. The 4B version works the same way; select the required version.
Deploy the service: Use SGLang (pip install "sglang==0.5.17") or vLLM to start the inference service, configuring --context-length 262144, --reasoning-parser qwen3, and --tool-call-parser qwen3_coder.
API calls: Use the OpenAI SDK with base_url set to the local service address, such as http://localhost:8000/v1, and send requests just as you would through the OpenAI API.
Resource considerations: The usable context length depends on GPU memory. If memory is insufficient, reduce the context length before deployment.

Agent-Native training: The training data comes directly from real execution trajectories rather than traditional question-and-answer data. The objective directly targets task completion instead of adapting a general-purpose model after the fact.
Utilization of failed trajectories: The training set retains both successful records and replaced failed records, teaching the model where errors are likely to occur and how to recover from them.
Routing signal reuse: Online C0–C3 capability-level signals are converted offline into training criteria, creating a closed loop of evaluation feedback → data allocation → model updates.
High performance at a small size: The 4B version’s Agent performance already surpasses that of multiple general-purpose models with more parameters. The 9B version achieves an average score of 69.04, exceeding its base model by 3.44 points, with lower deployment costs.
Long context: Natively supports 262,144 tokens and can be extended to approximately 1.01 million tokens, making it suitable for long-horizon Agent tasks.

GitHub repository: https://github.com/TokenRhythm/NeoHorse
Hugging Face model collection: https://huggingface.co/collections/TokenRhythm/neohorse-1
arXiv technical paper: https://arxiv.org/pdf/2609.08183

Agent task execution: Serves as an execution model in Harnesses such as OpenClaw, automatically completing real computer tasks such as file operations and software use, while recovering autonomously after failures.
Tool calling and API orchestration: Automatically selects, combines, and calls external APIs in multi-tool environments, making it suitable for building automated workflow assistants.
Deep search and research assistance: Actively retrieves and cross-validates information over multiple rounds and provides an evidence chain to support the generation of research reports with citations.
Enterprise office automation: Handles long-horizon, multi-step tasks such as schedule adjustments, data organization, and report generation in diverse workplace scenarios like WorkBuddy Bench.