Long-horizon coding agents have an uncomfortable cost structure: a single trajectory can run dozens or hundreds of steps, and the field's habit is to hand every step to a frontier cloud model. NVIDIA's research team has just posted EdgeWise on arXiv, and the paper opens by puncturing exactly that assumption — not every step needs a frontier model's capability, but every step is currently being billed at frontier-model prices.

Cloud-model value is concentrated in a few critical steps

The paper starts with an observation experiment. The team analyzed 2,400 coding-agent trajectories and bucketed each action into six categories — write, read, search, command, think, and other. They then routed a random 40% of steps to cloud GPT-5.5 or Claude Opus 4.8. The result is surprising but reasonable: even when cloud coverage is fixed at 40%, performance is essentially indistinguishable from full-cloud execution, sometimes even better. Combined with Pearson correlation analysis, the only step types that meaningfully track final-answer quality are writes and complex reads — where you ask the cloud matters more than how often you ask it.

That observation pins EdgeWise's design in place. A small local model handles most actions. The cloud gets woken up only at the few critical nodes — code writes and complex reads. Everything else is billed at local-compute prices.

Step-level RL teaches a 9B model when to call for help

EdgeWise ships two delegation strategies. The first is binary delegation: every step passes through an external judge model that decides whether the local SLM should keep going or hand the next step to the cloud. The judge is post-hoc routing — robust but you pay for two model calls per trajectory.

The second is the paper's centerpiece — self-delegation. Qwen3.5-9B is first supervised-fine-tuned, then trained with step-level reinforcement learning so that "when should I call the cloud" is baked into the model's weights. The reward comes from final task success, propagated back to each delegation decision via fine-grained credit assignment. During training, the model learns that calling too late costs you, calling too early wastes money. At inference time, no judge is needed — the local 9B model itself decides whether to forward the step.

The paper also proposes a third pattern: context management. When the cloud gets called, does the local model ship the full interaction history or only the current step? EdgeWise's answer is that full history usually wins — at the cost of a few more tokens. If you want to claw back more cost, you can filter out irrelevant steps.

The numbers: 53% / 63% / 50% across three pivotal cuts

The main table pairs Qwen3.5-9B locally with GPT-5.5 or Opus 4.8 in the cloud, capped at 80 turns per task. On SWE-ReBench, self-delegation cuts cloud cost by 53% (with GPT-5.5) and 63% (with Opus 4.8). On Terminal-Bench 2.1, the cuts are 35% and 33% respectively. Pass@3 stays comparable to all-cloud execution — the average delta is exactly 0.0%.

Cross-harness transfer is tested across four setups (training on Terminus2, evaluating on OpenHands, π, and mini-swe-agent). Self-delegation still cuts cost by at least 25% on unseen harnesses — evidence that "when to call the cloud" is a harness-agnostic meta-skill, not a framework-specific trick.

Finally, the real-world test: EdgeWise runs on a local DGX Spark workstation, cutting cost another ~50% while keeping Pass@3 on par with an A100 cloud setup. This is real-money deployment, not paper benchmarks — something a developer can actually use at the edge.

So what

EdgeWise gives a clean, executable answer to a question the field keeps debating without resolution: can a small model be the primary solver? Yes, but only if it knows which step it can't handle. Step-level RL internalizes "which step I can't handle" into the 9B model itself, so no external judge has to babysit every trajectory. It forces a re-examination of the folk wisdom that long-horizon agents must always run on frontier models — the cost structure can be redrawn, provided the local model learns to call for help at the right moment.

References: EdgeWise paper (arXiv:2610.17560, https://arxiv.org/abs/2610.17560); NVIDIA project page (https://research.nvidia.com/labs/lpr/edge-wise/).