zurichquants/OpenThinkerAgent-8B-RL
OpenThinkerAgent-8B-RL by OpenThoughts-Agent is an 8 billion parameter Qwen3-based model, fine-tuned with reinforcement learning for agentic coding tasks. It is designed to operate as a tool-using agent, capable of issuing shell commands and reasoning over terminal output to solve software engineering problems. With a context length of 40,960 tokens, this model excels in environments requiring interactive problem-solving and code generation. It represents the final RL-trained checkpoint in the OpenThoughts-Agent SFT→RL recipe, optimized for specific task distributions.
Loading preview...
OpenThinkerAgent-8B-RL Overview
OpenThinkerAgent-8B-RL is an 8 billion parameter model developed by OpenThoughts-Agent, representing the final, RL-trained checkpoint in their SFT→RL recipe. Built on the Qwen3-8B architecture, this model is specifically designed for agentic coding tasks, enabling it to function as a tool-using agent within environments like the terminus-2 harness. It can issue shell commands, perform edits, and reason over terminal output to solve software engineering problems.
Key Capabilities & Training
- Agentic Coding: Optimized for interactive problem-solving in software engineering contexts, leveraging external tools and environments.
- Reinforcement Learning: Undergoes on-policy reinforcement learning (RL step 45) using the OpenThoughts-Agent-RL-5K task set, building upon a cold-start SFT base.
- Architecture: Based on Qwen3-8B, featuring 36 layers, 4096 hidden size, and a 40,960-token context length (with RL rollouts using a 32,768-token serving window).
- Training Data: Initial SFT on OpenThoughts-Agent-SFT-ColdStartForRL-10K (9,437 task/trajectory pairs), followed by RL on 5,000
pymethods2test-largetasks.
Intended Use Cases
This model is ideal for applications requiring an agentic coding model that can interact with sandboxed environments to solve software engineering tasks. It inherits general capabilities from Qwen3-8B and specialized agentic behaviors from its SFT and RL stages. Users should note that outputs, including shell commands, may require review and should be executed in secure, sandboxed environments. The RL stage was optimized for pymethods2test/SWE-Smith-style tasks, which may affect generalization to other domains.