Orca-Agent-v0.1

In depth details behind the training, including the training code, is all open sourced here

Description

Orca-Agent-v0.1 is an orchestration agent that acts as the brain of the operation - it receives the user's task but never touches code directly. Instead, it:

• Analyses the task and breaks it into focused subtasks
• Dispatches explorer agents to understand the system
• Delegates implementation work to coder agents with precise instructions
• Verifies all changes through additional explorer agents
• Maintains the context store with all discovered knowledge

Performance

Orchestrator	Subagent	Terminal Bench
Qwen3-Coder-480B	Qwen3-Coder-480B	19.7%
Orca-Agent-v0.1-14B	Qwen3-Coder-30B	18.25%
Qwen3-14B	Qwen3-Coder-30B	7.0%

This places Orca-Agent-v0.1 (14B) + Qwen3-Coder-Flash (30B MoE), within striking distance of Qwen3-Coder-480B using the same architecture which placed #26 on TerminalBench when it was published recently in my other project.

Example model output

<think>
Okay, now with task_003 identifying a backup, the next step would be to attempt recovery using this identified file.
</think>

<task_create>
agent_type: 'coder'
title: 'Attempt recovery using the identified backup file'
description: |
  Execute the following steps to recover using the backup identified in task_003:
  1. Move the identified backup file (/tmp/terraform_work/.terraform.tfstate.tmp) to the target location (/infrastructure/recovered_state.json).
  2. Verify the operation by checking:
     - The existence of the new file at the target location
     - The size matches the original backup file
     - The file permissions are set appropriately (rw-r--r--)
max_turns: 10
context_bootstrap:
  - path: '/tmp/terraform_work/.terraform.tfstate.tmp'
    reason: 'The backup file identified in task_003'
context_refs:
  - 'task_003'
</task_create>

Model training Overview

• Full fine-tune of Qwen3-14b
• 32x H100s
  • 16x for training
  • 8x inference for Orca-Agent
  • 8x inference for subagent (Qwen3-Coder-30B-A3B)
• Trained with GRPO + curriculum learning
• Batch size 256, 64 rollouts per task
• More details here

Serving model

vLLM

vllm serve Danau5tin/Orca-Agent-v0.1

SGLang

python -m sglang.launch_server \
--model-path Danau5tin/Orca-Agent-v0.1

The agent's orchestration code can be found here.