Recursive Self-Improvement, Minus the Recursion

What it does NeoHorse-1 is a pair of post-trained causal language models—4B and 9B parameters, built on Qwen3.5—that target text-based agent tasks, tool use, coding, and instruction following. A routing harness assigns work across a model pool, records tool interactions and outcomes, estimates capability demand, and feeds that signal into the next training mixture. The current release closes one lap of an evaluation–selection–update loop; chaining multiple laps is the stated next step toward recursive self-improvement.

The interesting bit Most post-training treats data curation as a static preprocessing step. NeoHorse treats execution trajectories as a living training signal through routing-guided curriculum SFT and on-policy distillation, preserving harness context rather than discarding it. The project is essentially an attempt to turn an evaluation harness into a curriculum generator.

Key highlights

• Two Apache 2.0 checkpoints (4B and 9B) with native 262k context, extensible up to roughly 1M tokens
• Routing-guided curriculum SFT and on-policy distillation that keep harness context in the training signal
• Heavy data hygiene: exact/near-duplicate removal, evaluation decontamination, six-dimensional semantic scoring, and subscene-level Scene/Goal/Outcome labels
• GGUF, 8/5/4-bit quantized, and MLX variants for Apple silicon alongside standard BF16 weights
• Text-in/text-out interface for self-hosted inference; model cards carry per-size evaluation tables

Caveats

• Recursive self-improvement is described as a future goal, not an achieved outcome; the current harness completes one evaluation-selection-update cycle
• The README is upfront that extending the loop across successive iterations is the next step, so expect a research prototype rather than a finished product

Verdict Worth a look if you are building agent training infrastructure and want an open-weight experiment in closed-loop post-training. Skip it if you need a drop-in replacement for your current model stack.