Kimi K2: Open Agentic Intelligence (arXiv:2507.20534)

Fields

id

sources/model-cn-kimi-k2-paper

URL

https://arxiv.org/abs/2507.20534

Publisher

arXiv / Moonshot AI

Kind

whitepaper

Published

2025-07-28

Accessed

2026-10-03

notes

Primary source for Kimi K2's vendor-stated 1T total / 32B activated and for its architecture lineage: K2 is DeepSeek-V3-shaped (MLA, kv_lora_rank 512, 384 routed + 1 shared expert, 8 active per token) but trained with the Muon optimizer. Confirms the MLA cache geometry used for the kv_cache_bytes_per_token derivation.

Notes

Primary source for Kimi K2's vendor-stated 1T total / 32B activated and for its architecture lineage: K2 is DeepSeek-V3-shaped (MLA, kv_lora_rank 512, 384 routed + 1 shared expert, 8 active per token) but trained with the Muon optimizer. Confirms the MLA cache geometry used for the kv_cache_bytes_per_token derivation.

Cross-references

This record references no other records.