NeoHorse-1 tests a feedback loop for agent training

TokenRhythm has released NeoHorse-1, an open-weight model family designed for tool use, coding, instruction following, and text-based agent harnesses. The release includes 4B and 9B checkpoints post-trained from Qwen3.5, with weights available through Hugging Face and ModelScope under the Apache 2.0 license.

The project explores recursive self-improvement, or RSI, through an external training pipeline. An agent harness records model executions, identifies capability gaps, selects useful traces, and uses them to train a successor checkpoint. Each published checkpoint remains fixed during inference.

TokenRhythm provides BF16 safetensors and GGUF quantizations at 8-bit, 5-bit, and 4-bit precision. The GGUF releases support local inference through llama.cpp-compatible runtimes.

Agent traces become curriculum

The proposed system places a routing harness between users and a pool of models. It assigns each task to a model, records tool calls and outcomes, estimates which capabilities the task required, and sends that feedback into the next training mixture. Updated checkpoints can then return to the model pool, creating an evaluation-selection-update cycle.

1. The harness assigns a task and records tool calls, retries, corrections, and the final outcome.
2. The data pipeline removes unsuitable or contaminated traces and applies capability-level labels.
3. Selected trajectories feed post-training for the next checkpoint.
4. The updated checkpoint returns to the harness for another evaluation cycle.

The training recipe combines routing-guided curriculum supervised fine-tuning with routing-guided on-policy distillation. Supervised fine-tuning teaches from selected examples, while on-policy distillation trains on trajectories produced by models operating inside the current harness. Preserving tool responses, execution state, and harness context allows the model to learn from the sequence of decisions that produced an outcome.