Why this matters
Deployed agent systems produce rich execution traces (tool calls, outcomes, routing signals) that are rarely reused as training signal. This project demonstrates a concrete path: capture harness interactions, label capability demand, and turn those trajectories into a staged supervised/post-training curriculum that narrows the gap between smaller and larger models while preserving agentic context.
Key Capabilities
- Agentic post‑training pipeline: converts routed execution trajectories (interleaved reasoning, tool calls, harness context) into supervised fine‑tuning and on‑policy distillation, using routing signals to build a three‑stage curriculum.
- Practical model artifact: a ~4B causal LM (text-only weights, safetensors/BF16) derived from Qwen3.5‑4B and repackaged for long context inference (262,144 native; extensible further). Vision weights are not included.
- Measured gains: the post‑training protocol raises a ten‑benchmark macro average from 58.94 to 64.87 (+5.93), with notable improvements on agentic and coding benchmarks (e.g., HumanEval +9.75).
- Deployment examples: released with instructions for self‑hosting (SGLang, vLLM) and served via OpenAI‑compatible endpoints; emphasizes tool‑call parsers and reasoning parsers tuned for Qwen family semantics.
Who it's for & trade‑offs
Great fit if you are researching agentic systems, routing‑based curricula, or want a compact model tuned for tool use and coding where preserving execution context matters. It is useful for experimenting with recursive self‑improvement (RSI) prototypes and for self‑hosted inference with very long contexts.
Look elsewhere if you need a multimodal model (vision weights are omitted), a model with turnkey safety alignment for high‑risk production use, or the absolute top‑tier single‑model performance (NeoHorse‑1 narrows gaps with larger models but is not a 9B replacement). Expect GPU memory and FP/BF16 constraints for large context serving and that integrating the routing harness requires additional infrastructure and annotation tooling.
