Team-ACE/EnvACE-Qwen3-1.7B

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

EnvACE-Qwen3-1.7B is a 2 billion parameter tool-interactive agent model developed by Team-ACE, based on Qwen/Qwen3-1.7B with a 32768 token context length. It is trained using the EnvACE framework, which employs "world rehearsal" to internalize environment dynamics, allowing the model to act as its own environment. This model excels at long-horizon tool use by generating tool calls and then rehearsing the environment's response internally, making it suitable for agentic reinforcement learning tasks without external simulators.

Loading preview...

EnvACE-Qwen3-1.7B: A Tool-Interactive Agent Model

EnvACE-Qwen3-1.7B is a 2 billion parameter model from Team-ACE, built upon the Qwen/Qwen3-1.7B architecture. Its core innovation lies in the EnvACE framework, which enables long-horizon tool use through a novel training paradigm called world rehearsal. Instead of relying on external environments or simulators, the model internalizes environment dynamics within its own parameters, effectively acting as its own environment during training.

Key Capabilities and Training Innovations

  • World Rehearsal: The model alternates between generating a tool call and then simulating the environment's response internally, conditioning subsequent decisions on this rehearsed feedback.
  • Joint End-to-End Optimization: Both the 'acting' and 'rehearsal' roles share a single set of parameters, trained together using role-wise GRPO from task-success rewards, eliminating the need for external simulators.
  • Internalized World Model: Through this rehearsal process, the policy develops an internal world model that directly supports decision-making, allowing for private rehearsal before committed execution at test time.
  • Evaluated Performance: The method has been rigorously evaluated on benchmarks such as BFCL-v4, τ²-Bench, VitaBench, and FinMCP-Bench, demonstrating strong and transferable performance.

Use Cases

  • Agentic Reinforcement Learning: Ideal for scenarios requiring agents to interact with tools and environments over long horizons.
  • Tool Use and Function Calling: Designed to handle complex tool interaction sequences by internalizing the effects of actions.
  • Research in RL and Agent Systems: Provides a robust framework for exploring agent behavior without external environment dependencies.

For full integration and training with tool-interactive environments, developers can refer to the EnvACE repository.

Finetunes
1 models