Why this matters
Most model improvements are one‑shot: train, evaluate, repeat manually. This project demonstrates a prototype feedback loop that records routing decisions and tool executions from a heterogeneous agent harness, converts those execution traces into validated training examples, and uses capability‑guided allocation to shape subsequent post‑training — a practical step toward a harness‑mediated recursive self‑improvement workflow.
Key Capabilities
- Agent‑native post‑training: routing signals organize a curriculum of supervised fine‑tuning and routing‑guided on‑policy distillation so the model learns from real harness trajectories (reasoning, tool calls, context preserved).
- Measured gains: the 9B track reports a ten‑benchmark macro average of 69.04 (vs 65.60 for the Qwen3.5‑9B base), with notable improvements on agentic and coding benchmarks.
- Long context and deployment: native 262,144‑token context (extensible to ~1,010,000), weights provided as safetensors in BF16; examples and serving recipes for sglang and vLLM are included for self‑hosting.
- Data hygiene and labeling: pipeline includes duplicate removal, evaluation decontamination, structural validation, six‑dimensional semantic checks, and subscene Scene/Goal/Outcome labeling to convert executions into training signal.
Who it's for & trade‑offs
Great fit if you need an LLM tuned for agentic tool use, multi‑step execution traces, or coding assistants that can be integrated into a routing/harness infrastructure. The model is practical for research teams wanting a deployable checkpoint (OpenAI‑compatible serving examples are provided).
Look elsewhere if you require multimodal (vision) weights or an out‑of‑the‑box instruction‑tuned chat experience focused solely on conversational polish rather than harnessed agent execution. Also note this release contains text‑only weights derived from Qwen3.5‑9B; visual components are not included.
Where it fits
Positioned between base LLM checkpoints and full agent platforms: it’s a finetuned, agent‑aware model intended to improve real harness behavior through iterative post‑training. Compared with its Qwen3.5‑9B base it targets agentic/tooling gains rather than purely scaling or multimodal features.
Brief methodology note
The core mechanism is a routing harness that (1) routes turns to model/service tiers, (2) logs predicted capability demand, tool calls and outcomes, (3) validates and semantically labels resulting trajectories, and (4) uses capability‑guided allocation to form the next training mixture. This closes an evaluation→selection→update loop that can be iterated toward recursive self‑improvement.
