A 3-billion-parameter agent closes in on models ten times its size

AI is moving rapidly beyond static question-answering toward autonomous agents that can plan, reason and act across complex, multistep environments. Microsoft Research has released an open-source framework aimed at making research in that area cheaper: Orchard.

The shared-infrastructure idea

At the centre of the framework sits Orchard Env, a reusable environment service for training and evaluating agents across task domains.

The claim here is economic rather than technical. In agent research, every team builds its own environments, data pipelines and evaluation workflows from scratch — work that can take longer than the research itself. Having one infrastructure support software-engineering, web-navigation and personal-assistant agents at once removes that duplication.

More than that, agents can be trained directly inside real deployment harnesses such as Codex, OpenClaw and ZeroClaw. The distinction matters: the gap between the training environment and the running environment is one of the known reasons an agent that works in the lab stumbles in the field.

The number: 3 billion parameters

Three demonstration models ship with the framework: Orchard-SWE, Orchard-GUI and Orchard-Claw. What they share is showing that relatively small open-weight models can achieve strong results on complex real-world tasks.

The most striking measurement comes from Orchard-SWE. On SWE-bench Verified, which measures software-engineering tasks, the model reaches 69.7 percent. With value-model reranking that rises to 73.0 percent.

What makes the figure meaningful is its cost: the result comes from only about 3 billion active parameters, approaching frontier systems that use models more than ten times larger.

Why it matters

This is another instance of a finding that keeps recurring: model size alone is not decisive. When the task is narrowly defined and training happens inside the real operating environment, a far smaller model can close on a far larger one.

The practical consequence is cost. Running a model with three billion active parameters is a fraction of the cost of running one with thirty billion — and in always-on agent systems that difference repeats on every request.

Alongside the models and workflows, the project releases training data and evaluation methods, intended to give the broader research community a foundation for building and studying open agentic systems.

What releasing it openly means

The framework being open source means the result is verifiable. The SWE-bench Verified score can be tested independently, and because the training data and evaluation methods are published, how the result was reached can be inspected.

That places it in a different category from figures announced by closed systems. The gap between a claim resting on a company's own measurement and one anybody can reproduce determines how much the same percentage actually carries.