Team-ACE/EnvACE-Qwen3-8B

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

EnvACE-Qwen3-8B is an 8 billion parameter tool-interactive agent model developed by Team-ACE, based on the Qwen3-8B architecture. It is specifically trained with the EnvACE framework for long-horizon tool use, utilizing 'world rehearsal' to internalize environment dynamics rather than external interaction. This model excels at agentic reinforcement learning tasks, demonstrating strong performance on benchmarks like BFCL-v4, τ²-Bench, VitaBench, and FinMCP-Bench.

Loading preview...

EnvACE-Qwen3-8B: An Agentic Model for Long-Horizon Tool Use

EnvACE-Qwen3-8B is an 8 billion parameter model from Team-ACE, built upon the Qwen3-8B architecture, and specialized for tool-interactive agentic tasks. Its core innovation lies in the EnvACE framework, which employs a novel training approach called world rehearsal. Instead of relying on external environments, the model internalizes environment dynamics within its own parameters, effectively acting as its own environment.

Key Capabilities & Training Innovations

  • World Rehearsal: The model alternates between generating a tool call and then simulating the environment's response, conditioning subsequent decisions on this rehearsed feedback.
  • Joint End-to-End Optimization: Both the 'acting' and 'rehearsal' roles share a single set of parameters, optimized together using role-wise GRPO from task-success rewards, eliminating the need for external simulators.
  • Internalized World Model: Through rehearsal, the policy develops an internal world model, enabling private rehearsal before committed execution at test time for improved performance.

Performance & Benchmarks

Evaluated on challenging benchmarks such as BFCL-v4, τ²-Bench, VitaBench, and FinMCP-Bench, EnvACE-Qwen3-8B achieves an Overall score of 32.91 across the first three and TF1 46.78 on FinMCP-Bench. These results demonstrate strong and transferable performance, outperforming several other models in its class, including Simulator-8B, TOUCAN-7B, and EnvScaler-8B.

Good For

  • Developing tool-interactive agents that require long-horizon planning.
  • Research into agentic reinforcement learning and internal world models.
  • Applications where simulating environment dynamics internally is beneficial or external environments are costly/unavailable.
  • Tasks requiring complex decision-making and sequential tool use.

Finetunes
1 models