Kimi K3.1 Pro weights released as open-weight multimodal model

Moonshot AI has released the weights for Kimi K3.1 Pro on Hugging Face, and the model card was linked from the Hacker News front page on 2026-11-04. It matters because the release puts a very large, natively multimodal agentic model into the open-weight ecosystem, with published architecture details, deployment instructions and a broad set of benchmark comparisons that developers and researchers can inspect and test for themselves.

What was released?

According to the Hugging Face model card for moonshotai/Kimi-K3.1-Pro, Kimi K3.1 Pro is described as an open-weight, natively multimodal agentic model. The card presents it as the first entry in the K3.1 series and as post-trained from Kimi K3 using large-scale joint reinforcement learning across hundreds of downstream domains. The stated context length is 1,048,576 tokens, which the introduction rounds to a 1-million-token context window.

The card lists the supported modalities as text and image, and classifies the task as image-text-to-text. It also shows tags including Transformers, Safetensors, text-generation, conversational, custom_code, compressed-tensors and 8-bit precision. The licence shown on the page is listed as kimi-k3.1-pro. At the time captured in the source, the page showed 199 likes and a Moonshot AI organisation account with 20.4k followers, along with a single community item, but those engagement figures are a snapshot rather than a technical claim about the model itself.

For readers deciding whether to pay attention, the key point is not simply that another model appeared. The release includes the weights, a detailed model summary and a large evaluation table, which gives independent teams a basis for reproduction, comparison and further research rather than relying only on a product announcement. That transparency is central to why open-weight releases are followed closely by developers who build coding tools, agents and multimodal applications.

How big is it and how is it built?

According to the model card, Kimi K3.1 Pro is a Mixture-of-Experts model with 2.8 trillion total parameters and 104 billion activated parameters. It has 93 layers, including one dense layer. The attention design combines 69 KDA layers with 24 Gated DSA layers, with an attention hidden dimension of 7168 and 96 attention heads. The card says that, compared with Kimi K3, the full-attention Gated MLA layers have been replaced with Gated DeepSeek Sparse Attention layers, a change intended to support more efficient long-context modelling.

The sparse expert structure is unusually wide. The model card lists 896 experts, with 16 selected per token and 2 shared experts. Other listed dimensions include a latent MoE dimension of 3584, an MoE hidden dimension per expert of 3072, a vocabulary size of 160K and the SiTU-GLU activation function. For the sparse attention component, the DSA index head dimension is 128, the number of DSA index heads is 32, and the DSA index top-k value is 2048.

Vision is handled by a separate encoder. The card names MoonViT-V2 as the vision encoder, with 401 million parameters. Quantisation is listed as MXFP4 weights with MXFP8 activations, produced through quantisation-aware training. Those details matter in practice because total parameter count alone does not determine how a model behaves in deployment. Activated parameters, expert routing, attention design, vision encoding and quantisation all affect memory use, speed, accuracy and the hardware needed to run the system, and the card provides the figures needed to start that assessment.

What can it do, according to its benchmark results?

The model card says Kimi K3.1 Pro is intended to improve on coding, 3D creation, financial research, factual accuracy and other agentic capabilities. Its evaluation section compares the model against Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash-1017 and GLM-5.3. Unless otherwise specified, the card states that Kimi K3.1 Pro results use maximum reasoning effort, a temperature of 1.0 and top-p of 0.95. It also notes that a dash means the source table does not report a score, and that higher scores are better except where a benchmark explicitly says lower is better.

In coding, the card reports 65.7 on Terminal-Bench 4.0, 51.7 on SWE Marathon v1.1, 75.5 on DeepSWE v1.1, 72.8 on BenchCAD without tools and 92.7 on BenchCAD with tools. On its internal Kimi Code Bench 2.0 it reports 77.1, and on Kimi Code Bench 3.0 Long-horizon it reports 61.2. Those figures place it ahead of some comparison models on several coding tasks, while other rows in the same table show different leaders, which is a useful reminder that coding ability is not a single dimension.

In agentic and multimodal tasks, the card reports 70.0 on HLE-Full with tools, 92.1 on BrowseComp, 93.4 on CharXiv with python, 90.0 on MMMU-Pro with python and 95.4 on BabyVision with python. It reports 43.2 on Agents' Last Exam overall and 44.2 on the CLI subset. For reasoning and knowledge, it reports 94.2 on GPQA-Diamond and 89.0 on CharXiv without python. On the internal Kimi Hallu Bench Hard with tools, where lower is better, it reports 8.3, compared with 5.5 for GPT-6 Astra in the same table. The overall picture is therefore mixed but competitive, with strengths in several agentic, coding and multimodal rows and no claim of leading every category.

Where does it not lead?

A careful reading of the table also shows areas where Kimi K3.1 Pro does not lead the comparison set. On Terminal-Bench-Science 0.1, it reports 38.1, while GPT-6 Astra reports 68.1 and Claude Opus 5.5 reports 63.3. On Program Bench, it reports 31.5, compared with 65.0 for Claude Opus 5.5 and 53.5 for Claude Fable 5.1. On HLE-Diamond, it reports 39.4, while GPT-6 Astra reports 65.8 and Claude Opus 5.5 reports 60.9. In cybersecurity, it reports 74.7 on ExploitBench and 20.7 on ExploitGym, below the highest scores in those rows.

Those gaps do not cancel out the stronger results elsewhere, but they define the practical question for a reader. A team evaluating the model for terminal-based coding, document research, image understanding or long-horizon agent work should look at the specific benchmark that resembles its own task, check the evaluation footnotes for harness and tool differences, and then run its own tests. The card includes extensive footnotes explaining that some scores come from official leaderboards, some from model reports and system cards, and some from internal evaluations using harnesses such as Claude Code or Kimi Code. That context is essential because scores obtained with different tools, judges, hardware or fallback settings are not always directly interchangeable.

How can developers run it?

According to the Hugging Face page, developers can load the model with Transformers using an image-text-to-text pipeline or by loading the model directly with trust_remote_code enabled. The page also provides instructions for serving the model with vLLM and with SGLang, including example commands that expose an OpenAI-compatible chat completions endpoint on a local server. Docker options are shown as well, including Docker Model Runner and a Docker image for SGLang that maps GPU devices and the Hugging Face cache into the container.

The examples on the page use a text prompt asking for a one-sentence description of an image, paired with an image URL. That example is illustrative of the model's image-text-to-text use, but it is not a benchmark result. Developers who want to reproduce the published numbers will need to pay close attention to the evaluation settings in the card, including maximum reasoning effort and the sampling parameters, as well as the specific harness used for each benchmark. The page also links to inference providers, HuggingChat, notebooks on Google Colab and Kaggle, and local apps, giving several routes for initial experimentation before committing to a full deployment.

How does this relate to Grafy?

For Grafy readers, the most direct connection is coding and agent work. Grafy Code is an AI coding studio, so developers exploring Kimi K3.1 Pro for repository tasks, terminal workflows or long-horizon coding agents can use that workspace to compare model behaviour on their own code while reading the benchmark footnotes in the source. Because the model is multimodal and designed for agentic use, Grafy Chat offers a practical way to test prompting, image understanding and multi-model comparison side by side without first setting up local serving infrastructure.

The long context window also makes the release relevant to research workflows. Grafy Reader lets readers search and question PDFs and papers with AI, which is a useful counterpart when evaluating whether a 1-million-token context is genuinely helpful for a particular corpus or whether retrieval, chunking and citation checks still matter more. Grafy does not claim to host Kimi K3.1 Pro as a named feature in the list of apps provided for this article, so the sensible approach is to treat the Grafy workspace as a place to test workflows and compare outputs, while using the Hugging Face instructions for direct deployment of the weights.

Frequently asked questions

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

Yes. According to the Hugging Face model card, Kimi K3.1 Pro is presented as an open-weight model, and the page provides downloadable weights and deployment instructions. The card lists the licence as kimi-k3.1-pro, so anyone planning commercial or redistributive use should read that licence carefully before deploying it.

Does Kimi K3.1 Pro support image input?

Yes. The model card lists its modalities as text and image and classifies the task as image-text-to-text. It names MoonViT-V2 as the vision encoder, with 401 million parameters. The Hugging Face examples also show an image-text-to-text pipeline and prompts that combine text with an image URL.

What context length does Kimi K3.1 Pro have?

The model card lists a context length of 1,048,576 tokens and describes the model in its introduction as having a 1-million-token context window. That long context is one of the model's main design points, and the card says the Gated DSA attention change is intended to support more efficient long-context modelling.

Source: Hugging Face — moonshotai/Kimi-K3.1-Pro

---

Source: Hacker News — front page — Kimi K3.1 Pro model weights are out