keras/qwen_agent_world_35b_a3b

Model Overview

Qwen-AgentWorld

Qwen-AgentWorld is a native language world model developed by Qwen (Alibaba Cloud) designed for general agentic environment simulation. Unlike post-hoc environment adapters, Qwen-AgentWorld is trained from Continual Pre-Training (CPT) onward with environment modeling as its primary objective. Given an agent's action and interaction history, Qwen-AgentWorld uses long chain-of-thought reasoning to predict the next environment state across text and GUI interfaces.

Model Highlights

• Seven Unified Domains: Simulates 7 diverse environment domains within a single model: Model Context Protocol (MCP), Search, Terminal/CLI, Software Engineering (SWE), Android, Web, and OS.
• Native World Model: Built as an environment simulator from the CPT stage onward, followed by Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL via GSPO).
• Hybrid MoE Architecture: Combines Gated DeltaNet linear attention (3:1 ratio with Gated Attention) and sparse Mixture-of-Experts (256 total experts, 8 routed + 1 shared active per layer) for fast, memory-efficient inference.
• Agent Foundation Capabilities: Zero-shot generalization to out-of-domain environments (e.g., OpenClaw) and support for multi-turn tool-calling environments.
• Long Context: Native support for up to 262,144 tokens (256K context window).

More Details

• [Qwen AgentWorld Quickstart Notebook](coming soon..!)
• Qwen AgentWorld API Documentation
• Qwen AgentWorld Model Card
• KerasHub Beginner Guide
• KerasHub Model Publishing Guide

Installation

Keras and KerasHub can be installed with:

pip install -U -q keras-hub
pip install -U -q keras

Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instructions on installing them in another environment see the Keras Getting Started page.

---

Model Presets

Preset Name	Total Params	Active Params	Description	Architecture
qwen_agentworld_35b_a3b	35.1B	~3B	35B total parameter Qwen-AgentWorld MoE environment simulator model (~3B active parameters per token). Trained across 7 agentic domains.	Qwen3.5 MoE

---