Latest

Training on internalized skill workflows lets a 35B open-weights model beat far larger proprietary models on agentic benchmarks, shifting the edge from raw model scale to curated environment data. Builders running Claude Code or Codex harnesses can reuse the released environments and trajectories for their own SFT or RL pipelines. The results give Qwen-based deployments a concrete benchmark point against Claude Sonnet 4.6 and DeepSeek V4 Pro.

The key insight is architectural: letting an LLM handle semantic meta-reasoning while a GA handles dense numerical search removes the manual configuration bottleneck in controller tuning. Control engineers and robotics practitioners can apply the pattern to their own optimization pipelines without expensive bilevel search. Running on a compact memory buffer and a low-cost model at roughly $0.002 per run also makes the approach practical for iterative lab or industrial deployment.

The key finding is that skills can be internalized as model weights rather than injected at inference time, since the fine-tuned agent beats skill-assisted base models even without the skills provided. Agent engineers and fine-tuning practitioners can reuse the released environments and trajectories for SFT or RL with outcome-based rewards. The result also shows a 35B open Qwen-based model outscoring much larger closed models on SkillsBench, a useful cost comparison for teams building coding and tool-use agents.

The core result is that a quantized local model can match a cloud API on Vietnamese compliance extraction, removing the data-sovereignty objection that blocks LLM use under privacy law. Compliance teams and legal-tech builders in Vietnam face a 1 January 2026 deadline, and this shows a single 24 GB GPU suffices for the workload. The unmeasured value-level precision and mid-range token coverage are caveats to weigh before deployment.