DeepSeek-TNG R1T2 Chimera
TNG Technology Consulting · 2025-07-02 · 684.5B parameters
DeepSeek-TNG-R1T2-Chimera is TNG Technology Consulting's 671B-class Tri-Mind Assembly-of-Experts merge using DeepSeek-R1, DeepSeek-R1-0528, and DeepSeek-V3-0324 components, released under MIT. Its FP8 repository stores 684,531,386,000 tensor entries including quantization-scale tensors, while the published active size is a rounded 37B. The configuration sets a 163,840-token YaRN-scaled limit over an original 4,096-token RoPE window.
Benchmark scores
| Benchmark | Score |
|---|---|
| SnakeBench Average Apples | 1.3 |
| SnakeBench Best Apples | 8.0 |
| SnakeBench Rating | 16.6 |
| SnakeBench Total Apples | 19.0 |
| SnakeBench Win Rate | 50.0 |