48B total / 3B active per token (nominal MoE); 49,122,681,728 checkpoint parameters
1,048,576 tokens (1M)
MIT (explicit Base/Instruct checkpoint declarations; copyright/permission notices must be preserved; standalone HF LICENSE files absent)
Text-generation checkpoint using a layerwise hybrid of Kimi Delta Attention (KDA) and full Multi-Head Latent Attention (MLA), not Kimi K2 or a parallel concatenated attention branch. KDA extends Gated DeltaNet with channel-wise decay gating and a hardware-efficient chunkwise delta-rule algorithm using specialized diagonal-plus-low-rank transitions. The advertised 3:1 KDA/MLA pattern is implemented as 20 KDA and 7 MLA layers across 27 decoder layers; MLA uses NoPE. The first feed-forward layer is dense; subsequent MoE layers select 8 of 256 routed experts plus 1 shared expert. Released weights are predominantly BF16 with FP32 KDA gate parameters and no exported MTP layers. Pretrained on 5.7T tokens sampled from the K2 corpus with MuonClip, a WSD schedule and initial 4,096-token sequences, followed by annealing and long-context activation; both model card and launch config support 1M context without a YaRN override. The report describes multi-stage general/reasoning SFT and math/code/STEM RL with concurrent PTX/SFT, truncated importance sampling and adaptive KL/update scheduling; precise final-checkpoint stage budgets are not published. The shipped chat template has no thinking-mode switch, although reasoning benchmarks and RL training are reported. Publisher-reported Instruct results include RULER 94.8 at 1M, MMLU-Pro 72.7 and LiveCodeBench v6 pass@1 45.7. Controlled 1.4T-token same-backbone comparisons favor Kimi Linear on average, but it trails full MLA on LongBench V2, Frames and commit-level Long Code Arena; these ablations are not the released 5.7T checkpoint. Up to 75% less KV cache and approximately 6x faster decoding at 1M are publisher-specific comparison claims, not universal hardware guarantees: the report distinguishes larger-batch headline TPOT from batch-one tests and does not disclose complete hardware/precision settings in v1. An official eligible Base checkpoint also exists, but is excluded from this provisional Instruct-only card until its first public-download date and same-day grouping are resolved.