AIsaAIsaStudio 中文 Open AIsa Studio
PickHugging Face 热门模型· Models·

Moonshot AI releases Kimi K3.1 Pro, an open-weight multimodal agentic model with 2.8T total parameters and 1M context

Original title: moonshotai/Kimi-K3.1-Pro (image-text-to-text)

Moonshot AI released Kimi K3.1 Pro, an open-weight, natively multimodal agentic model and its most capable 3T-class model to date, post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. The model uses a MoE architecture with 2.8T total and 104B activated parameters, 896 experts with 16 selected per token, and a 1M-token context window; the team reports substantial improvements in coding, 3D creation, financial research, factual accuracy, and agentic capabilities.

Why it matters

The official README lays out K3.1 Pro's MoE architecture details and 1M context, letting readers gauge deployment requirements and capability positioning.

Read the original ↗AI Daily, Nov 5

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.

Analysis

What happened

Moonshot AI has released the open-weight model Kimi-K3.1-Pro on Hugging Face, where it currently ranks 6th among trending models with 479 likes and 797 recent downloads. The model's tags identify it as an image-text-to-text, conversational text-generation model.

The official introduction describes Kimi K3.1 Pro as an "open-weight, natively multimodal agentic model" and the team's most capable 3T-class model to date. It was post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, is the first model in the K3.1 series, and offers a 1-million-token context window.

Key details

  • The architecture is a Mixture-of-Experts (MoE) with 2.8T total parameters and 104B activated parameters per token.
  • It has 93 layers, of which 1 is dense; the attention layers consist of 69 KDA layers plus 24 Gated DSA layers, with an attention hidden dimension of 7168 and 96 attention heads.
  • DSA index specifics: index head dimension 128, 32 index heads, and a Top-K of 2048.
  • MoE configuration: 896 experts with 16 selected per token, plus 2 shared experts; latent MoE dimension 3584 and per-expert MoE hidden dimension 3072.
  • Vocabulary size is 160K, and the context length is 1 million tokens (the README table is truncated at this field; the introduction paragraph states the figure).
  • The model is distributed via the transformers library using safetensors and compressed-tensors (with 8-bit tags), includes custom_code, and uses a proprietary license named "kimi-k3.1-pro" under the "other" category.
  • The release includes a Tech Blog and an API Quickstart, and the model is also available on ModelScope.

Why it matters

  • With 2.8T total parameters, 896 experts, and 104B activated per token, it sits at the top tier of open-weight model scale; the compressed-tensors and 8-bit tags indicate official deployment-oriented compressed formats.
  • The 1-million-token context window and the hybrid KDA/DSA attention design (69 + 24 layers) underpin its long-context capability, and the sparse indexing mechanism with a Top-K of 2048 is notable.
  • The team claims substantial improvements in coding, 3D creation, financial research, factual accuracy, and other agentic capabilities. Positioned as a multimodal agentic model, it takes image and text input and produces text, targeting agent scenarios.

Who it affects

  • Developers and researchers building long-context, multimodal, agentic applications can obtain the weights via Hugging Face or ModelScope, or use the API directly.
  • The license is a proprietary "other" type, so the LICENSE file should be reviewed before commercial use; the model includes custom_code, which requires security and compatibility checks at deployment.
  • Practitioners tracking the open-weight LLM landscape should note that this is the first model in the K3.1 series, suggesting further releases in the series.
Original article

Hugging Face 热门模型第 6 名：moonshotai/Kimi-K3.1-Pro，479 赞，近期下载 797 次。 标签：transformers, safetensors, kimi_k31, text-generation, compressed-tensors, conversational, image-text-to-text, custom_code, license:other, eval-results, endpoints_compatible, 8-bit

README：

📰  Tech Blog |     🚀  API Quickstart

1. Model Introduction

Kimi K3.1 Pro is an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date. Post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains, it is the first model in the K3.1 series. With a 1-million-token context window, Kimi K3.1 Pro delivers substantial improvements in coding, 3D creation, financial research, factual accuracy, and other agentic capabilities.

2. Model Summary

Architecture Mixture-of-Experts (MoE)

Total Parameters 2.8T

Activated Parameters 104B

Number of Layers 93

Number of Dense Layers 1

Attention-Layer Composition 69 KDA + 24 Gated DSA*

Attention Hidden Dimension 7168

Number of Attention Heads 96

DSA Index Head Dimension 128

DSA Index Number of Heads 32

DSA Index Top-K 2048

Latent MoE Dimension 3584

MoE Hidden Dimension (per Expert) 3072

Number of Experts 896

Selected Experts per Token 16

Number of Shared Experts 2

Vocabulary Size 160K

Context Length</s

The original belongs to Hugging Face 热门模型 and its authors and is reproduced here for reading. To have it removed, write to support@aisa.one.