The race in agentic AI has a mantra: more parameters, more power. Yet a team at Microsoft Research has just broken this equation with Orchard, an open-source framework that lets software agents run at near-frontier levels using models with only 3 billion active parameters. The secret isn’t an exotic optimization trick, but a radical redesign of what those agents stand on: the execution environment.

The idea at the heart of Orchard is as simple as it is disruptive. Instead of embedding training and evaluation infrastructure inside a single framework, Orchard Env provides a lightweight, shared Kubernetes service that spins up isolated containers at scale. The same service – without modification – powers code-writing agents (Orchard‑SWE), web-navigation agents (Orchard‑GUI), and personal-assistant agents (Orchard‑Claw). The result is a reusable environment for data distillation, supervised fine-tuning, and reinforcement learning, slashing costs and eliminating dependence on proprietary sandboxes or opaque cloud services.

The numbers tell the story. Orchard‑SWE reaches 69.7% on SWE‑bench Verified, the strictest benchmark for automatic code repair. Using a 35-billion-parameter mixture-of-experts model with only 3 billion parameters active per token, the system climbs to 73% with a simple reranking step based on a value model. These results approach systems with more than 10 times the active parameters, all without resorting to hundreds-of-billions-parameter giants. Similarly, Orchard‑GUI – a 4-billion-parameter vision‑language model – averages 68.4% across three web-navigation benchmarks, ranking among the top open agents and matching proprietary solutions from OpenAI and Google.

Yet the real breakthrough isn’t the scores, but the architecture that made them possible. Orchard shifts the center of gravity from the model to the infrastructure. The shared Kubernetes environment enables training agents directly inside real deployment harnesses – Codex, OpenClaw, ZeroClaw – instead of simplified surrogates. A lightweight proxy records harness model calls as training data, while each rollout runs in its own container. This eliminates the mismatch between training and production, and learns from partial attempts rather than discarding them.

This choice has structural implications far beyond the lab. When open, reusable infrastructure allows compact models to match or approach the performance of mega-models, the cost per token of an agent plummets. You no longer need to rent cloud GPUs to serve enormous models; on-prem clusters sized for inference of a few billion active parameters are enough, with clear TCO and data sovereignty advantages. Companies can keep agents within their own boundaries, without exposing code and sensitive information to third-party services.

Orchard makes the prospect of commoditized agentic AI concrete, where the true competitive differentiator shifts to the execution environment and data pipelines, not model size. For those evaluating on-prem LLM deployment today, the framework provides a blueprint: a lightweight Kubernetes layer, data-efficient training (just 400 demonstrations guide Orchard‑GUI), and the ability to accumulate experience between training runs, distilling trajectories into reusable value models. It’s a paradigm shift that erodes the moat of cloud vendors and returns control over the entire agent lifecycle to organizations.

It’s no coincidence that Orchard was born inside Microsoft Research but released completely open source. The message is clear: the future of agents won’t be written in the weights of a single model, but in the ability to orchestrate flexible, retrainable, and sovereign environments. And that ability is now within reach of anyone with a Kubernetes cluster and the will not to outsource intelligence.