gordonhu/Tev-27B

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 5, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Tev-27B is a 27 billion parameter long-horizon terminal agent model developed by gordonhu, fine-tuned from Qwen3.8-27B using reinforcement learning with progress-potential reward shaping (PPRS) on the Tev-Gym environment. It excels at complex, multi-step computing and math tasks, demonstrating significant improvements in zero-shot transfer to longer horizons and maintaining strong mid-horizon coding performance. This model is designed for research into advanced agentic capabilities in terminal environments.

Loading preview...

Overview

Tev-27B is a 27 billion parameter terminal agent model, fine-tuned from Qwen3.8-27B. It specializes in long-horizon tasks, addressing common challenges like sparse training data and rewards through its unique training methodology. The model was developed using reinforcement learning on Tev-Gym, a novel environment synthesis approach, and employs progress-potential reward shaping (PPRS) for dense reward signals.

Key Capabilities

  • Enhanced Long-Horizon Task Solving: Achieves a 65% improvement in Pass@5 on the ALE-CM benchmark (0.17 to 0.28) compared to its base model, matching frontier models like GPT-6 Astra and Claude Opus 5.
  • Strong Zero-Shot Transfer: Demonstrates significant gains on unseen longer-horizon tasks, with Pass@1 increasing by 56% on Terminal-Bench 3.0 "CM" subset and 2.5x on Terminal-Bench 4.0 "CM" subset.
  • Maintained Mid-Horizon Coding: Achieves a Pass@1 of 0.794 on Terminal-Bench 2.1, on par with Claude Opus 4.8, indicating robust performance on coding tasks.
  • Trained on Diverse Environments: Utilizes Tev-Gym, which includes 1,170 verified long-horizon environments across 7 domains (software engineering, data engineering, AI/CS research, mathematics, security, infrastructure, quantum computing).

Good for

  • Research in Long-Horizon Agents: Ideal for academic and non-commercial research into AI agents capable of solving complex, multi-step problems in terminal environments.
  • Computing and Math Workflows: Excels in tasks requiring sequential execution of commands, problem-solving, and logical reasoning within a bash shell.
  • Benchmarking Agent Performance: Useful for evaluating agent capabilities on challenging, long-duration tasks where traditional models struggle due to sparse rewards.

Finetunes
1 models