2026-09-11 10:44

Introduction: Execution trajectory-driven training, routing signal reuse loop, failure recovery learning included. What is NeoHorse-1? NeoHorse-1 is a primitive rhythm joint venture with NoWen Xincore

What is NeoHorse-1?

NeoHorse-1 is an open-source Agent model jointly developed by Primitive Rhythm, Wuwen Xincore, Tsinghua University, Peking University, Hong Kong Chinese University, Alibaba, and others, available in 4B and 9B variants. The model employs Agent-Native post-training using real execution trajectories from Routing Harness as training data, reusing routing signals as the basis for curriculum ordering, enabling a closed-loop process of evaluation feedback → data proportioning → model update. It covers tool invocation, task planning, deep search, and code generation. The 9B version achieves an average score of 69.04, outperforming its base model by 3.44 points, natively supports up to 260,000 tokens context, and can be deployed locally.

Main Features of NeoHorse-1

• Tool Invocation: Automatically selects and invokes external tools based on task requirements to perform multi-step operations.
• Task Planning: Decomposes complex tasks into executable sub-steps, enabling long-horizon planning and scheduling.
• Deep Search: Proactively retrieves and integrates information across multiple interaction rounds to support evidence chains for conclusions.
• Code Generation: Generates usable code and iteratively debugs and fixes it based on environmental feedback until the task is completed.

Technical Principles of NeoHorse-1

• Harness Trajectory-Driven Agent-Native Training: Training corpus is centered on structured execution trajectories generated by real-world Routing Harness operations, encompassing requests, routing decisions, tool calls, environment feedback, and error recovery—adding three dimensions beyond traditional Q&A data: capability demands, execution decisions, and environmental outcomes. Both successful and failed trajectories are preserved, filtered via structural validation and six-dimensional quality assessment, directly aligning training objectives with task completion.
• Capability-Guided Data Proportioning: The router categorizes tasks into four capability tiers (C0–C3), repurposing routing signals offline as training guidance: routing records are split into “prediction—action—result” segments, where prediction segments define curriculum order and result segments identify capability gaps, enabling dynamic adjustment of data distribution in subsequent training cycles, forming an evaluation feedback → data proportioning → model update loop.
• Single-Round Validation via Recursive Self-Improvement (RSI): Comprises two components: Data-RSI refers to continuous model execution generating new trajectories, which are screened and naturally accumulated as future training material; Model-RSI involves updating the model based on evaluation shortcomings and reintegrating it into the Harness model pool. Currently, only single-cycle closure is validated, with signal and reward design still human-defined.

Follow WeChat and reply "open source" to join the AI Open Source Project Discussion Group

How to Use NeoHorse-1

• Install Download Tools: Execute pip install -U modelscope to install the ModelScope command-line tool.
• Download Model: Run modelscope download --model TokenRhythm/NeoHorse-1-9B --local_dir ./NeoHorse-1-9B to pull weights (the 4B version follows the same procedure, select as needed).
• Deploy Service: Launch inference service using SGLang (pip install "sglang==0.5.17") or vLLM, configuring --context-length 262144, --reasoning-parser qwen3, and --tool-call-parser qwen3_coder.
• API Invocation: Use OpenAI SDK with base_url pointing to your local service address (e.g., http://localhost:8000/v1), allowing requests just like calling OpenAI endpoints.
• Note Resource Requirements: Actual usable context length depends on GPU memory; reduce context length before deployment if memory is insufficient.

Core Advantages of NeoHorse-1

• Agent-Native Training: Training data originates directly from real execution trajectories rather than conventional Q&A corpora, targeting task completion instead of post-hoc adaptation of general-purpose models.
• Leveraging Failure Trajectories: Training sets preserve both successful and replaced failed records, enabling the model to learn “where errors occur and how to recover after failure.”
• Routing Signal Reuse: Offline transformation of online routing C0–C3 capability tier signals into training input enables a closed-loop process: evaluation feedback → data proportioning → model update.
• High Performance with Small Size: The 4B Agent version already surpasses several larger-parameter general models; the 9B version scores 69.04 (3.44 points above base), offering lower deployment costs.
• Ultra-Long Context: Native support for 262,144 tokens, extendable to approximately 1 million tokens, ideal for long-horizon Agent tasks.

Project Links for NeoHorse-1

• GitHub Repository: https://github.com/TokenRhythm/NeoHorse
• HuggingFace Model Hub: https://huggingface.co/collections/TokenRhythm/neohorse-1
• arXiv Technical Paper: https://arxiv.org/pdf/2609.08183

Competitive Comparison with Similar Projects

Application Scenarios of NeoHorse-1

• Intelligent Agent Task Execution: Acts as the execution model within systems like OpenClaw, automatically completing real computer tasks such as file operations and software usage, and autonomously recovering after failures.
• Tool Invocation & API Orchestration: Automatically selects, composes, and invokes external APIs in multi-tool environments, ideal for building automated workflow assistants.
• Deep Search & Research Assistant: Conducts multi-round proactive retrieval, cross-verification of information, and provides evidence chains to support report generation with citations.
• Enterprise Office Automation: Handles long-horizon, multi-step tasks across diverse workplace scenarios (e.g., WorkBuddy Bench), including schedule adjustments, data organization, and report generation.

Source: AI Tools Collection