NeoHorse-1-4B: A Small Model Testing the Road to Self-Improving AI
NeoHorse-1-4B fine-tunes Qwen3.5-4B with a routing harness for agentic tasks, scoring 64.87 average, up 5.93 points over its base model.
What is NeoHorse-1-4B?
NeoHorse-1-4B is an open-weight 4-billion-parameter language model released by TokenRhythm, built by post-training Qwen3.5-4B for agentic tasks: tool use, coding, and instruction following. What sets it apart isn’t size, it’s the training method. TokenRhythm calls it a “routing harness,” a system that assigns tasks across a pool of models, logs what happens, and feeds that feedback back into training. The company frames NeoHorse-1-4B as an early prototype on the way to recursive self-improvement (RSI), not a finished RSI system.
TL;DR
- NeoHorse-1-4B is a fine-tune of Qwen3.5-4B released by TokenRhythm under Apache 2.0, aimed at agentic use cases like tool calling, coding, and following complex instructions.
- The model uses a routing harness that sends tasks to a mix of models, records tool interactions and outcomes, and uses that data to decide what to train on next.
- TokenRhythm frames this as a prototype loop toward recursive self-improvement, where updated models re-enter the harness for another round of evaluation and selection, though the current release only demonstrates one such loop.
- On a ten-benchmark suite, NeoHorse-1-4B posts a 64.87 macro average, a 5.93-point gain over base Qwen3.5-4B’s 58.94.
- It shows the largest jumps on agentic benchmarks like WorkBuddy Bench (+9.79) and VitaBench (+10.50), with smaller gains on function-calling tests like BFCL v4 (+0.77).
- The model ships as BF16 safetensors, natively supports 262,144 tokens of context, and is extensible up to roughly 1,010,000 tokens.
- Training data went through deduplication, decontamination, and structural validation, plus a six-dimensional semantic evaluation and scene-level labeling of task trajectories.
How does the routing harness actually work?
The routing harness sits between raw training data and the model itself. Instead of training on a fixed, pre-baked dataset, TokenRhythm routes incoming tasks to a heterogeneous pool of models, capturing how each one handles tool calls, multi-step plans, and task outcomes. That trace data (what tools were called, whether the task succeeded, where the model struggled) gets used to estimate which capabilities are underdeveloped across the pool. Those capability gaps then shape the next round of training data.
The research behind NeoHorse-1-4B describes two techniques layered on top of this: routing-guided curriculum SFT, which sequences training examples based on where the harness detects need, and routing-guided on-policy distillation, which turns actual execution trajectories into training signal rather than relying purely on static demonstrations. The stated goal is to preserve the surrounding execution and harness context around each response, not just the final answer, so the model learns from the process, not just the output.
Why call it a step toward recursive self-improvement?
Recursive self-improvement describes a system that can evaluate its own weaknesses, generate better training data or better versions of itself, and repeat that cycle without a human redesigning the process each time. NeoHorse-1-4B doesn’t claim to do this autonomously. What TokenRhythm has built is a loop: route tasks, measure capability gaps, retrain, and let the updated model re-enter the harness for another round of evaluation.
That’s a meaningful distinction. A single iteration of “train on your own weak points” is standard post-training practice at this point. What TokenRhythm is pitching as novel is the infrastructure to make that loop repeatable and self-directed over successive generations, with the routing harness acting as a standing evaluator rather than a one-off benchmark run. The model card is explicit that extending this loop across iterations is “the next step,” meaning NeoHorse-1-4B represents one pass through the cycle, not proof that the cycle compounds indefinitely.
What do the benchmark numbers actually show?
TokenRhythm compared NeoHorse-1-4B against five other open-weight models in the same size class: the Qwen3.5-4B base model, Gemma-4-E4B-it, Nanbeige-4.2-3B, Agents-A1-4B, and Spark-X2.5-4B. Across ten benchmarks, NeoHorse-1-4B posted a 64.87 macro average versus 58.94 for its base model, a gain of 5.93 points.
The gains aren’t uniform. On agent-specific benchmarks, the improvements are large:
- QwenClawBench: 44.68 vs. 38.47 base (+6.21), the best score among all six models tested.
- WorkBuddy Bench: 34.41 vs. 24.62 base (+9.79), also the top score.
- PinchBench: 77.33 vs. 71.19 base (+6.14), again the best result.
- VitaBench: 32.00 vs. 21.50 base (+10.50), the largest single delta, though not the top score overall (Agents-A1-4B scored 39.25).
On more established function-calling benchmarks, the gains shrink. BFCL v4 moved from 61.02 to 61.79 (+0.77), landing behind Nanbeige-4.2-3B’s 67.28. tau2-Bench also saw a modest change, with NeoHorse-1-4B trailing Nanbeige-4.2-3B’s 85.08. The pattern suggests the routing harness approach pays off most clearly on newer, more agent-flavored benchmarks that reward multi-step tool orchestration, and less on tasks that are already well-covered by existing fine-tuning approaches.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Is NeoHorse-1-4B worth using for agentic or coding work?
For teams building tool-using agents or coding assistants and constrained to small, self-hostable models, NeoHorse-1-4B is a reasonable candidate to evaluate. It’s Apache 2.0 licensed, ships in standard safetensors/BF16 format, and its context window (262K tokens natively, extensible toward 1M) is generous for a 4B model, useful for agents that need to hold long tool-call histories or large codebases in context.
The tradeoffs are the ones you’d expect from any 4B model: it won’t match larger models on raw reasoning depth, and its benchmark wins are concentrated in agentic and tool-use scenarios rather than across the board. On standard function-calling evaluations like BFCL v4, it’s competitive but not dominant against similarly sized peers. Anyone evaluating it should treat the benchmark table as a guide to where the model specifically improved (tool orchestration, multi-step agent tasks) rather than assume uniform gains over the Qwen3.5-4B base.
How was the training data prepared?
TokenRhythm describes a fairly involved data pipeline behind the routing harness. It includes exact and near-duplicate removal to avoid overrepresenting similar examples, evaluation decontamination to prevent benchmark test data from leaking into training, and structural validation to catch malformed examples. Beyond that, the pipeline applies a six-dimensional semantic evaluation and labels training examples at the subscene level with Scene, Goal, and Outcome tags, effectively annotating each fragment of an agent trajectory with what was happening and whether it worked.
This level of curation matters more for agentic post-training than for typical instruction tuning, since agent trajectories are long, multi-step, and easy to contaminate with either duplicated tool-call patterns or leaked benchmark scenarios. It’s also consistent with the routing harness’s purpose: if the harness is going to convert live task traces into training signal, the traces need to be clean and well-labeled before they’re pooled with other data.
Frequently Asked Questions
What base model is NeoHorse-1-4B built on?
It’s a fine-tune of Qwen3.5-4B, using the same roughly 4B parameter architecture but repackaged for text-only inference (vision components from the base model are not included).
What license is NeoHorse-1-4B released under?
Apache 2.0, making it free to use, modify, and redistribute, including for commercial purposes.
What does “routing harness” mean in this context?
It refers to a system that distributes tasks across multiple models, logs tool use and outcomes, and uses that data to guide what the next round of training should focus on, forming the basis of TokenRhythm’s post-training loop.
Does NeoHorse-1-4B actually achieve recursive self-improvement?
No. TokenRhythm describes it as an early prototype demonstrating one iteration of an evaluate-select-update loop. Extending that loop across multiple generations is described as future work, not something this release proves out.
How big is NeoHorse-1-4B’s context window?
It supports 262,144 tokens natively and can be extended up to approximately 1,010,000 tokens, according to the model card.
