Kimi Linear: An Expressive, Efficient Attention Architecture (arXiv:2510.26692)

Fields

id

sources/kimi-linear-tech-report

URL

https://arxiv.org/abs/2510.26692

Kind

paper

Accessed

2026-10-19

notes

Read https://arxiv.org/abs/2510.26692 for the benchmark figures on models/kimi-linear-48b-a3b. As recorded by the reader: Values are the released Kimi-Linear-Instruct row of the tech report's instruct-model table (48B total / 3B active, 5.7T tokens): MMLU-Pro (EM) 72.7, GPQA-Diamond (Avg@8) 71.7, AIME 2025 (Avg@64) 58.6, MATH500 (Acc.) 94.6, LiveCodeBench v6 (Pass@1) 45.7. The HF card's headline figure is MMLU-Pro 51.0, but that is the 1.4T-token architecture-comparison number (also Figure 1 / Table 4), not the released checkpoint; 72.7 is the released instruct model. 'Humaneval +' 70.9 omitted because it is HumanEval+ (different benchmark).

Notes

Read https://arxiv.org/abs/2510.26692 for the benchmark figures on models/kimi-linear-48b-a3b. As recorded by the reader: Values are the released Kimi-Linear-Instruct row of the tech report's instruct-model table (48B total / 3B active, 5.7T tokens): MMLU-Pro (EM) 72.7, GPQA-Diamond (Avg@8) 71.7, AIME 2025 (Avg@64) 58.6, MATH500 (Acc.) 94.6, LiveCodeBench v6 (Pass@1) 45.7. The HF card's headline figure is MMLU-Pro 51.0, but that is the 1.4T-token architecture-comparison number (also Figure 1 / Table 4), not the released checkpoint; 72.7 is the released instruct model. 'Humaneval +' 70.9 omitted because it is HumanEval+ (different benchmark).

Cross-references

This record references no other records.