DeepSeek-TNG-R1T2-Chimera
10 months ago
- #DeepSeek
- #AI
- #Chimera
- DeepSeek-TNG-R1T2-Chimera is a new model, successor to the original DeepSeek R1T Chimera.
- It is constructed using the Assembly of Experts method with three parent models: R1-0528, R1, and V3-0324.
- Fixes the <think> token consistency issue present in the original R1T Chimera.
- Operates at a new sweet spot in intelligence vs. output token length, being 20% faster than R1 and more intelligent in benchmarks like GPQA and AIME-24.
- Recommended as a drop-in replacement for R1, a cheaper alternative to R1-0528, and generally preferred over R1T unless specific traits of R1T are needed.
- Not recommended for function-calling intensive applications due to influence from the R1 parent model.
- Benchmarking now includes AIME24, AIME25, and GPQA-Diamond, showing larger score differences between R1 and R1T Chimera.
- Architecture is a DeepSeek-MoE transformer-based language model, released on 2025-07-02.
- Users in the EU are advised to comply with the EU AI Act guidelines effective August 2nd, 2025, or cease using the model.
- Feedback is encouraged via email or X.com, and citation details are provided for academic use.