DeepSeek-R1T-Chimera
TNG Technology Consulting · 2025-04-27 · 684.5B parameters
DeepSeek-R1T-Chimera is TNG Technology Consulting's open-weight Assembly-of-Experts merge, combining DeepSeek-V3-0324's non-routed weights with DeepSeek-R1's routed-expert weights without fine-tuning or distillation. Its released FP8 checkpoint preserves the DeepSeek-V3 architecture's 684,489,845,504 model-weight parameters; the repository additionally stores quantization-scale tensors, while TNG describes the main architecture with rounded labels of 671B total and 37B active. Its 61-layer MLA configuration uses partial RoPE with YaRN scaling from an original 4,096-token window to 163,840 positions.
Benchmark scores
| Benchmark | Score |
|---|---|
| Longform Creative Writing | 59.4 |
| SnakeBench Average Apples | 3.8 |
| SnakeBench Best Apples | 12.0 |
| SnakeBench Rating | 26.0 |
| SnakeBench Total Apples | 64.0 |
| SnakeBench Win Rate | 60.0 |