- Tertia Lux
- Posts
- Everyone’s chasing bigger models.
Everyone’s chasing bigger models.
Then a German consulting firm stitched together a faster one — on their internal R&D days.
James D. Wilson
July 05, 2025
No fine-tuning.
No retraining.
Just Assembly-of-Experts.
TNG Technology Consulting's R1T2 Chimera merges DeepSeek R1, R1-0528, and V3-0324 at the tensor level.
It inherits structure from R1. Concision from V3. Reasoning from 0528.
It’s not smarter.
It’s cheaper.
200% faster output than 0528 — mostly due to 60% fewer tokens.
Benchmark scores sit ~10% below 0528.
About 10% higher than R1.
So: slightly less accurate.
Much shorter.
Roughly 3x cheaper to run.
Released MIT-licensed.
Free to use.
Also: doesn’t support function calling.
Not ideal for tools or agents.
May not generalize well beyond reasoning benchmarks.
The speed gain doesn’t come from architecture.
It comes from verbosity cuts.
Not all teams will make that trade.
It’s a clever remix.
Not a miracle.
But for cost-sensitive enterprise inference?
It might be the new baseline.
Paper: arXiv:2506.14794
#deepseek #ai #llm