← All models

Qwen-AgentWorld-35B-A3B

Published Publisher Alibaba

Overview

Qwen-AgentWorld-35B-A3B is an open-weight language world model designed to simulate environments for AI agents. Given an agent’s action and interaction history, it predicts the next environment observation using reasoning. It uses a mixture-of-experts architecture with 35 billion total parameters, 3 billion active parameters, and a 256K-token context window. (huggingface.co)

Capabilities and uses

The model covers seven agent-interaction domains: MCP, search, terminal, software engineering, Android, web, and operating systems. It can be used to simulate environment responses, develop and evaluate agents, and research agent training. Alibaba also released AgentWorldBench, a benchmark covering these seven domains. (github.com)

Deployment and limitations

The weights are available under the Apache 2.0 license and can be deployed with frameworks including Transformers, SGLang, and vLLM. The official repository notes that the checkpoint contains language-model weights only, despite visual component definitions in its architecture; its vLLM example therefore uses the --language-model-only flag. Simulated observations are predictions rather than verified results from a real environment and should not be treated as evidence that an action actually occurred. (huggingface.co)

Pricing

As of November 16, 2026, input and output API prices for this exact model could not be verified in either Alibaba Cloud’s mainland China CNY pricing documentation or its international USD pricing documentation. No token rate or billing unit is confirmed for this model. Self-deployment may incur infrastructure costs; pricing for similarly named Qwen models should not be assumed to apply. (help.aliyun.com)

Release and official sources