Production model

Kimi K1.5

Moonshot AI · Released 2025-01-22 · Not publicly disclosed parameters · 128,000 ctx · closed weights

Architecture

Positional encoding	—
Norm placement	—
Norm type	—
QK-Norm	—
Activation	—
Attention	—
MoE	—
Other	Multimodal (text + image inputs), Long-context training: up to 128K, with reported strong results at long needle-in-haystack tasks, Architecture details (parameter count, attention recipe, MoE/dense, normalization choices) not publicly disclosed in the tech report

Kimi K1.5 (January 2025) is Moonshot AI’s reasoning-focused release, predating the open-weights Kimi K2 by about six months. The paper “Kimi k1.5: Scaling Reinforcement Learning with LLMs” (arXiv 2501.12599) reports the model’s training recipe in detail — long-context training, multimodal input handling, and a custom RL post-training stack that scaled to long chain-of-thought reasoning behaviors competitive with frontier closed models at the time of release.

What the paper does not report at the level needed for an architecture entry:

Total or active parameter counts
Specific attention recipe (MHA / GQA / MLA / other)
Normalization placement and type
Activation function inside the FFN
Whether the model is dense or MoE

Per the citation policy, those architecture fields are left null rather than guessed. The model is included for navigability — readers arriving from “what does Kimi K1.5 use?” should land here and learn that the architecture is not publicly documented at the paper’s level of detail.

Disclosure level is documented: paper available, weights closed at the time of release. Kimi K2 (July 2025) is the open-weights successor in the Kimi family; that entry has full architecture detail because Moonshot released the model card and inference code alongside the weights.

Sources

https://arxiv.org/abs/2501.12599

Export

BibTeX

@article{arxiv_2501_12599,
  title         = {Kimi K1.5},
  author        = {Moonshot AI},
  year          = {2025},
  eprint        = {2501.12599},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2501.12599}
}

CSL JSON

{
  "id": "arxiv_2501_12599",
  "type": "article-journal",
  "title": "Kimi K1.5",
  "author": [
    {
      "literal": "Moonshot AI"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2025
      ]
    ]
  },
  "URL": "https://arxiv.org/abs/2501.12599",
  "number": "2501.12599",
  "source": "arXiv"
}

RIS

TY  - JOUR
TI  - Kimi K1.5
AU  - Moonshot AI
PY  - 2025
JO  - arXiv
AN  - arXiv:2501.12599
UR  - https://arxiv.org/abs/2501.12599
ER  -