News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgfnKimi K3.1 Pro

Kimi K3.1 Pro

Open-weight multimodal agentic model with 1M token context

Moonshot AI released Kimi K3.1 Pro, an open-weight multimodal model designed for complex agent tasks. It features a one-million-token context window and uses sparse attention to handle long documents efficiently.

Share on WhatsApp

  • Features a 1-million-token context window
  • Uses Gated DeepSeek Sparse Attention layers
  • Part of the 3T-parameter class models
  • Post-trained with joint reinforcement learning
  • Open-weight multimodal agentic model

What it is

Kimi K3.1 Pro is a large language model developed by Moonshot AI. It belongs to the K3.1 series and represents the company’s most capable system to date, operating in the three-trillion-parameter class. The model is natively multimodal and designed for agentic workflows, meaning it can plan and execute multi-step tasks. It supports a context window of one million tokens, allowing it to process vast amounts of information at once. This capacity enables significant improvements in coding, three-dimensional creation, and financial research compared to its predecessor.

How it came about

The model was post-trained from Kimi K3 using large-scale joint reinforcement learning. This training process covered hundreds of downstream domains to enhance its practical utility. Kimi K3.1 Pro serves as the first release in the K3.1 series. It builds upon the foundation of Kimi K3 but introduces architectural changes to improve efficiency and capability in long-context scenarios. The development focused on enhancing factual accuracy and agentic skills through extensive reinforcement learning across diverse fields.

Inside the model

The architecture uses a Mixture of Experts design with MXFP4 weights and MXFP8 activations. A key change from Kimi K3 is the replacement of full-attention Gated MLA layers with Gated DeepSeek Sparse Attention layers. This modification aims to make long-context modeling more efficient. The model undergoes quantization-aware training to maintain performance while reducing computational demands. These structural adjustments allow the system to handle the one-million-token context window effectively while managing the complexity of a three-trillion-parameter scale.

What it can do

The model shows strong performance in coding benchmarks such as Terminal-Bench and SWE-Marathon. It excels in agentic tasks evaluated by benchmarks like Agents' Last Exam and OSWorld V2.1. In multimodal tests, it processes PDFs and images for tasks in GDP.pdf and MMMU-Pro. Internal evaluations highlight improvements in financial research and three-dimensional creation. The system also demonstrates reduced hallucination rates on internal benchmarks when using tools. It performs competitively against other leading models in reasoning and knowledge assessments.

How to use it

Kimi K3.1 Pro is available as an open-weight model. Users can access the model weights directly, allowing for local deployment or further fine-tuning. The release includes technical details for integration via API or local infrastructure. This open approach enables researchers and developers to inspect the model structure and adapt it for specific applications. The availability of weights distinguishes it from closed proprietary systems that only offer API access.

Questions people ask about Kimi K3.1 Pro

What is the context window size of Kimi K3.1 Pro?

The model supports a context window of one million tokens. This allows it to process and reason over extremely large documents or codebases in a single pass without losing track of earlier information.

Is Kimi K3.1 Pro an open-source model?

It is an open-weight model. This means the model weights are publicly available for download and use, though it may not meet all criteria for strict open-source definitions depending on license terms.

What architectural change distinguishes it from Kimi K3?

It replaces the full-attention Gated MLA layers found in Kimi K3 with Gated DeepSeek Sparse Attention layers. This change improves efficiency when handling long contexts.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Kimi K3.1 Pro yet, so the facts on this page come from the sources listed above: huggingface.co/moonshotai/Kimi-K3.1-Pro. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.