DeepSeek R1T2 Chimera: Faster, High‑Performance LLM with Assembly of Experts

DeepSeek R1T2 Chimera, a newly released open‑source large language model, combines the DeepSeek R1‑0528, R1 and V3‑0324 base models using the Assembly of Experts (AoE) technique.

According to benchmarks, it runs 200% faster than R1‑0528 and 20% faster than R1, while achieving higher scores than R1 on GPQA‑Diamond and AIME‑24, though it does not yet match the top‑tier R1‑0528.

The model adopts a DeepSeek‑MoE Transformer architecture with 671 billion parameters and is released under the MIT license on Hugging Face.

Compared with previous variants, R1T2 offers a “Tri‑Mind” fusion, improving think‑token consistency and providing a better balance between intelligence and output length.

Compared to DeepSeek R1, R1T2 can serve as a drop‑in replacement with superior performance.

Compared to R1‑0528, it is more cost‑effective when the highest intelligence is not required.

Compared to the earlier R1T Chimera, R1T2 is recommended unless a specific personality or token‑consistency issue is critical.

Compared to DeepSeek V3‑0324, V3 is faster but R1T2 delivers stronger reasoning capabilities.

Limitations include slightly lower performance than R1‑0528 on the most difficult benchmarks, higher reserved token latency, lack of function‑calling support, and recent changes in evaluation metrics.

For detailed technical information, see the paper “Assembly of Experts: Linear‑time construction of the Chimera LLM variants with emergent and adaptable behaviors” (https://arxiv.org/pdf/2506.14794).