Tev-27B
A long-horizon agent model trained on the Tev framework. It exhibits cross-domain generalization capability and performs well on multiple agent benchmarks.
Timeline (last 90 days)
The model trained with the PPRS method is Tev-27B.
Tev-27B achieves Claude Opus 4.8-level performance on the medium-horizon task TB 2.1.
Tev-27B, trained only on ALE-related data, can nonetheless generalize to Terminal-Bench, suggesting that it has learned some domain-agnostic long-horizon task execution capability rather than merely memorizing specific task solutions.
Tev simultaneously provides the data infrastructure (Tev-Gym), the training method (PPRS), and the validation results (Tev-27B), forming a complete closed loop.
Without having seen Terminal-Bench training data, Tev-27B surpassed GLM 5.2 and Gemini 3.7 Flash on the TB 3.0 and 4.0 subsets, and reached the performance of Claude Opus 4.8 on TB 2.1.
Tev-27B's generalization ability is limited to similar domains (computation, mathematics); transferring to entirely new domains (legal research, hardware design) would most likely require retraining.
Tev-27B was trained entirely with PPRS on the computational and mathematical sciences domains of ALE (Agents' Last Exam), matching GPT-6 Astra and Claude Opus 5 on the Pass@5 metric of ALE-CM.
All Facts (7)
Tev simultaneously provides the data infrastructure (Tev-Gym), the training method (PPRS), and the validation results (Tev-27B), forming a complete closed loop.
50%UnverifiedTev-27B achieves Claude Opus 4.8-level performance on the medium-horizon task TB 2.1.
50%UnverifiedTev-27B, trained only on ALE-related data, can nonetheless generalize to Terminal-Bench, suggesting that it has learned some domain-agnostic long-horizon task execution capability rather than merely memorizing specific task solutions.
50%UnverifiedWithout having seen Terminal-Bench training data, Tev-27B surpassed GLM 5.2 and Gemini 3.7 Flash on the TB 3.0 and 4.0 subsets, and reached the performance of Claude Opus 4.8 on TB 2.1.
50%UnverifiedTev-27B's generalization ability is limited to similar domains (computation, mathematics); transferring to entirely new domains (legal research, hardware design) would most likely require retraining.
50%UnverifiedTev-27B was trained entirely with PPRS on the computational and mathematical sciences domains of ALE (Agents' Last Exam), matching GPT-6 Astra and Claude Opus 5 on the Pass@5 metric of ALE-CM.
50%UnverifiedThe model trained with the PPRS method is Tev-27B.
50%