Tsinghua SIGS Robot Lab has released VeriLoop E2, a 27B post-trained model built on Qwen3.8-27B for verifiable code, mathematics, and scientific reasoning. The release introduces VeriLoop-Governed Recurrence (VGR), which separates generation from external verification to ensure state transitions strictly improve evidence dimensions.
- Post-trained on 1,841,831 records spanning software engineering, math, and science.
- Achieves scores including 76.2% on SWE-bench Pro, 98.3% on AIME 2026, and 93.94% on GPQA Diamond.
- Releases a full GGUF precision ladder from BF16 down to IQ1_M, with all tiers derived directly from the canonical BF16 checkpoint.
- IQ1_M variant reaches 16.79 GiB with only +0.3191% PPL drift and supports MTP speculative decoding.
- Includes a reproducible certificate for the critical-line zero proportion of the Riemann ζ function.
The release provides open weights, evaluation evidence, and technical reports to allow independent verification of quantization retention and downstream performance.