Open-Source · Agentic Modeling

🌳 Orchard

An Open-Source Agentic Modeling Framework

* First Authors  ·  † Second Authors

Performance comparison scatter plots. Left: Orchard-SWE (35B-A3B) reaches 73.0% on SWE-bench Verified with value-model reranking (69.7% with RPR-based RL), approaching frontier mixture-of-experts systems 10–30× larger. Right: Orchard-GUI (4B) reaches 68.4% average success across WebVoyager, Online-Mind2Web, and DeepShop.
Performance comparison. Left: Orchard-SWE (35B-A3B) reaches 73.0% on SWE-bench Verified with value-model reranking (69.7% with RPR-based RL), approaching frontier MoE systems 10–30× larger. Right: Orchard-GUI (4B) achieves 68.4% average success across WebVoyager, Online-Mind2Web, and DeepShop, making it the strongest open-source GUI agent while staying on par with proprietary systems from OpenAI and Google.

Abstract

Agentic modeling aims to transform large language models (LLMs) into autonomous agents that can solve complex tasks through planning, reasoning, tool use, and multi-turn interaction with external environments. We present Orchard, an open-source framework for scalable agentic modeling. At its core is Orchard Env, a thin, Kubernetes-native environment service that provides reusable primitives for sandbox lifecycle management, designed to operate across task domains, agent harnesses, and different training stages. On top of Orchard Env, we build three agentic modeling recipes. Orchard-SWE targets software-engineering agents: we introduce credit-assignment SFT to learn from productive segments of unresolved trajectories, and for RL go beyond sparse outcome rewards with a progression of signals — Balanced Adaptive Rollout (BAR) for sample-efficient sparse-reward optimization, plus two denser rewards, on-policy distillation (OPD) and rubric-based process reward (RPR). We further introduce historical experience distillation, which compresses rollouts from past experiments into a small value model that reranks candidates at inference time. On the Qwen3.5-35B-A3B backbone it reaches 69.7% with RPR-based RL and 73.0% with best-of-N value-model reranking, setting a new state of the art among open-source methods while approaching frontier systems more than 10× larger. Orchard-GUI trains a 4B vision-language computer-use agent using only 0.4K distilled trajectories and 2.2K open-ended tasks, achieving 74.1%, 67.0%, and 64.0% success rates on WebVoyager, Online-Mind2Web, and DeepShop, respectively (68.4% average) — the strongest open-source model while remaining competitive with proprietary systems from OpenAI and Google Gemini. Orchard-Claw targets personal assistant agents for productivity workflows such as email, calendar, and daily tool use: trained with only 0.2K synthetic tasks, it achieves 59.6% pass@3 on Claw-Eval and 73.9% when paired with a stronger ZeroClaw harness. Collectively, these results show that a thin, open, harness-agnostic environment layer enables the reuse of agentic data, training recipes, and evaluation protocols across domains and harnesses.

How Orchard fits together

One thin environment layer powers every stage of the agent lifecycle — data generation, supervised fine-tuning, reinforcement learning, and evaluation — across three task domains.

⌨️

Orchard-SWE

Coding agents on real repositories

SWE-bench Verified
SWE-bench Multilingual
🖥️

Orchard-GUI

VLM-based computer-use agents

WebVoyager DeepShop
Online-Mind2Web
🤖

Orchard-Claw

Personal assistant agents

Claw-Eval
harness-agnostic API · reusable across recipes
🌳

Orchard Env

A thin, Kubernetes-native environment service — reusable primitives for sandbox lifecycle management

create / destroy exec (async) apply patch file I/O screenshot & action network policy TTL cleanup
schedules isolated, reproducible sandboxes
Kubernetes / AKS

Per-sandbox namespaces, NetworkPolicy isolation, autoscaling pools

Browser sandbox

Containerized browser driven via Playwright — screenshots & full web interaction

Local / Docker

Lightweight containers for fast iteration and CI

Three recipes, one environment

Orchard-SWE

Coding agents

Orchard-SWE curates 107K trajectories distilled from MiniMax-M2.5 and Qwen3.5-397B, then adds credit-assignment SFT to salvage productive segments of unresolved trajectories and a progression of denser RL signals — Balanced Adaptive Rollout (BAR), on-policy distillation (OPD), and rubric-based process reward (RPR) — plus best-of-N reranking with a small value model.

  • SFT + RL (RPR)69.7%
  • + value model73.0%
  • BackboneQwen3.5-35B-A3B

New state of the art on SWE-bench Verified among open-source models of comparable size.

Orchard-GUI

Computer-use agents

A compact 4B vision-language agent trained on only 0.4K distilled trajectories plus 2.2K open-ended tasks, driving a live browser through screenshots and actions — the strongest open-source computer-use model while staying competitive with proprietary systems.

  • WebVoyager74.1%
  • Online-Mind2Web67.0%
  • DeepShop64.0%

Outperforms prior open-source agents and even its 235B teacher model, on a tiny data footprint.

Orchard-Claw

Personal assistant agents

Trained with just 0.2K synthetic tasks, Orchard-Claw shows that the same environment layer transfers to assistant-style agents — and improves further when paired with a stronger ZeroClaw harness at inference time.

  • Claw-Eval (pass@3)59.6%
  • + ZeroClaw harness73.9%
  • Training tasks0.2K

Harness-agnostic: swap the harness without changing the Env.

Key training ingredients

01

Trajectory distillation

Large teacher models (MiniMax-M2.5, Qwen3.5-397B) generate agent trajectories that are distilled into smaller, deployable student models — 107K trajectories for SWE alone, but as few as hundreds for GUI and Claw.

02

Credit-assignment SFT

Instead of discarding failed rollouts, Orchard uses retrospective value estimation to extract the productive segments of unresolved trajectories — converting partial progress that standard outcome-filtered SFT throws away into supervision.

03

Balanced Adaptive Rollout (BAR)

A sample-efficient RL rollout strategy for sparse rewards: it adaptively assembles reward-balanced trajectory groups so exploration is spent where the outcome signal is most informative across tasks of varying difficulty.

04

Denser RL rewards: OPD & RPR

Two dense signals go beyond sparse outcome rewards — on-policy distillation (OPD) supplies a token-level teacher signal at lower rollout cost, and rubric-based process reward (RPR) scores verification quality along the whole trajectory, reaching 69.7% on SWE-bench Verified.

05

Historical experience distillation

Rollouts accumulated across past RL runs are compressed into a small value model that reranks candidates at inference time (best-of-N), lifting Orchard-SWE to 73.0% on SWE-bench Verified.

Inside Orchard Env

The environment service exposes a small, uniform API for managing isolated sandboxes. The same client code drives a Kubernetes-backed coding sandbox or a Playwright-driven browser for GUI agents — so data generation, training, and evaluation all share one interface.

Sandbox lifecycle

Create, query, and tear down isolated sandboxes with per-sandbox namespaces, custom CPU/memory, and automatic TTL cleanup.

Async execution

Submit commands and poll for results; per-sandbox serialization with global concurrency control keeps multi-turn rollouts consistent.

Patches & files

Apply git patches and upload/download/list files — the primitives agent harnesses need to read, edit, and test code.

Browser control

For GUI agents, a containerized browser driven via Playwright adds screenshots and full web interaction — clicks, typing, scrolling, and navigation.

Isolation by default

Calico NetworkPolicy blocks egress by default; each sandbox runs in its own namespace with resource requests equal to limits.

Scales out, cheaply

Runtime agent injection and direct Pod-IP routing hit 0.28 s average exec latency and sustain a 1,000-sandbox stress test at 100% success — at ~2× lower cost than E2B and Daytona (10× with spot).

orchard_env — quick start 
# One client, the same API for every recipe
from aks_modal import SandboxClient

with SandboxClient() as client:
    with client.create_sandbox("python:3.11-slim") as sandbox:
        # run a command and read the result
        result = sandbox.exec("pytest -q", timeout=600)
        print(result.stdout, result.exit_code)

        # apply an agent-proposed patch, then re-test
        sandbox.apply_patch(patch_text)
        sandbox.exec("python -m pytest tests/")

Results at a glance

RecipeModelBenchmarkScore
SWE Qwen3-30B-A3B-Thinking SWE-bench Verified (SFT)64.3%
SWE-bench Verified (SFT+RL)67.5%
Qwen3.5-35B-A3B SWE-bench Verified (RPR-RL)69.7%
SWE-bench Verified (+ value model)73.0%
SWE-bench Multilingual (RPR-RL)62.3%
GUI Qwen3-VL-4B-Thinking WebVoyager74.1%
Online-Mind2Web67.0%
DeepShop64.0%
Claw Qwen3-30B-A3B-Thinking Claw-Eval (pass@3)59.6%
Claw-Eval + ZeroClaw (pass@3)73.9%

All models are open-source and state-of-the-art among comparable-size open models on their respective benchmarks.

BibTeX

cite 
@article{peng2026orchard,
  title   = {Orchard: An Open-Source Agentic Modeling Framework},
  author  = {Peng, Baolin and Yao, Wenlin and Wu, Qianhui and Cheng, Hao and
             Yu, Xiao and Yang, Rui and Ge, Tao and Sordoni, Alessandro and
             Yuan, Xingdi and Shen, Yelong and He, Pengcheng and Zhang, Tong and
             Yu, Zhou and Gao, Jianfeng},
  journal = {arXiv preprint arXiv:2605.15040},
  year    = {2026}
}