Kimi-K3.1-Pro 1M-Context Multimodal Agent Model: ~1128GB Memory

At a Glance
| Item | Value |
|---|---|
| Repository | moonshotai/Kimi-K3.1-Pro |
| Family guide | Kimi-K3.1 guide (1 articles) |
| Publisher guide | Moonshot AI: models and licenses |
| Published | 2026-11-02 |
| License | kimi-k3.1-pro |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Moonshot AI has released “Kimi K3.1 Pro," an open-weight native multimodal agent model. This is the company’s most powerful 3T-class model released to date. Based on Kimi K3, it has undergone post-training through large-scale collaborative reinforcement learning spanning hundreds of downstream domains. Furthermore, compared to the conventional Kimi K3, it achieves more efficient long-context modeling by replacing the full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers. Featuring a 1 million token context window, it delivers significant improvements in coding, 3D creation, financial research, factual accuracy, and other agent capabilities.
Specifications
- Architecture: Mixture-of-Experts (MoE) (Total parameters: 2.8T, Active parameters: 104B)
- Context Length: 1048576
Performance
According to measurements by the publishers, performance on major benchmarks is as follows (comparisons have been narrowed down).
→ Scroll horizontally to see all columns
| Benchmark | Kimi K3.1 Pro (max) | GPT-6 Astra (max, w/ fallback) | Claude Opus 5.5 (max, w/ fallback) | DeepSeek-V4.1-Flash-1017 (max) |
|---|---|---|---|---|
| Coding | ||||
| Terminal-Bench 4.0 | 65.7 | 58.2 | 66.4 | 42.3 |
| SWE Marathon v1.1 | 51.7 | 42.5 | 46.3 | 40.0 |
| Program Bench | 31.5 | 50.0 | 65.0 | 29.1 |
| Kimi Code Bench 2.0 (Internal) | 77.1 | — | — | 76.1 |
| Agentic | ||||
| HLE-Full w/ tools | 70.0 | 57.2 | 67.7 | 58.8 |
| BrowseComp | 92.1 | 91.5 | 90.8 | — |
| Agents’ Last Exam (ALL) | 43.2 | 41.8 | 42.6 | 37.7 |
| Automation Bench-AA | 73.3 | 68.5 | 69.5 | 66.6 |
| OSWorld V2.1 | 79.6 | 75.8 | 81.8 | 69.3 |
| Reasoning, Knowledge & Multimodal | ||||
| GDP.pdf | 34.6 | 34.2 | 30.6 | 17.2 |
| HLE-Full | 51.1 | 53.6 | 59.8 | 42.6 |
| HLE-Diamond | 39.4 | 65.8 | 60.9 | 21.6 |
| GPQA-Diamond | 94.2 | 96.1 | 93.9 | 92.2 |
| MMMU-Pro | 84.1 | 86.9 | 87.7 | 80.6 |
| Cyber Security | ||||
| ExploitBench | 74.7 | 100 | 91.0 | — |
According to measurements by the publishers, the model records a very high score of 94.2 on GPQA-Diamond, which tests difficult scientific questions, making it extremely strong in terms of specialized knowledge. It also shows high performance with 92.1 on BrowseComp, an agent-related task. On the other hand, it scores 31.5 on Program Bench in the coding domain, which is an inferior result in this metric compared to models like the comparison target Claude Opus 5.5 (65.0). On Terminal-Bench 4.0, which measures agent (terminal operation) capabilities, it records 65.7, putting it at a high level close to Claude Opus 5.5 (66.4).
Strengths and Use Cases
Kimi K3.1 Pro is a native multimodal agent model that handles both text and images. Leveraging its extremely long context window of 1 million tokens, it demonstrates suitability for advanced tasks such as the following:
- Agent Capabilities: Significant improvements have been made in coding, 3D creation, financial research, and various agent tasks. Specifically, it exhibits high performance in BrowseComp involving browser operations, OSWorld V2.1 involving OS GUI operations, and Automation Bench-AA which measures automation tasks.
- Multimodal Understanding: It excels at understanding combining images and text, recording high scores on benchmarks such as PDF document analysis (GDP.pdf), graph and chart analysis (CharXiv, ChartoGraphy), and MMMU-Pro which tests image understanding.
- Advanced Reasoning and Specialized Knowledge: It demonstrates extremely high accuracy on GPQA-Diamond, which handles difficult scientific questions, making it suitable for reasoning tasks requiring specialized knowledge.
- Cyber Security: Based on benchmark results like ExploitBench, utilization is also anticipated for tasks related to vulnerability discovery and exploits.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| More than 1128GB of VRAM (multi-GPU or CPU offload required) | Q2_K | 940.0GB | 1128.0GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build DevQuasar/moonshotai.Kimi-K3.1-Pro-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but DevQuasar/moonshotai.Kimi-K3.1-Pro-GGUF provides a GGUF build you can use.
License — kimi-k3.1-pro: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
We have not confirmed that this site meets the commercial-use terms of this model’s license (kimi-k3.1-pro). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
How to Get It
- Distribution format: safetensors
- Recommended inference engines:
- vLLM – SGLang – TokenSpeed
The Kimi K3.1 Pro API is available by selecting kimi-k3.1-pro at https://platform.kimi.ai. APIs compatible with OpenAI and Anthropic are also provided.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-11 | nvidia | NVFP4 | nvidia/Kimi-K3.1-Pro-NVFP4 | – |
| 2026-11-17 | DevQuasar | GGUF | DevQuasar/moonshotai.Kimi-K3.1-Pro-GGUF | Q2_K 1128.0GB (does not fit a single consumer GPU) |
File sizes of each build:
- Available builds in DevQuasar/moonshotai.Kimi-K3.1-Pro-GGUF: Q2_K 940.0GB / MXFP4 1405.3GB
In addition, 21 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- Ling-3.0-flash-Sante-fp8 Text Generation Model: ~144GB Memory
- Ling-3.1-flash-1019 Text Generation Model: ~641GB Memory, GGUF Builds
- Nanbeige4.3-33B 33B Model for Coding and Agents: 12GB+ VRAM
- Xing4.0-29B-A4B-w8a8 Released: W8A8 Quantized MoE Model
What to Read Next
- Find models by VRAM (This model does not fit even in 80GB; its smallest build needs about 1128GB) → VRAM quick reference, including models over 80GB
- Explore the same model family → Kimi-K3.1 family overview (1 articles, 2 converted builds)
- Engines that run this model → llama.cpp / Ollama / vLLM
- What Q2_K mean and where to get this model → GGUF format guide and models / NVFP4 / MXFP4 format guide and models / Quantization and model-format glossary
- Learn about the publisher → Moonshot AI: models, licenses and articles
Sources
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: DevQuasar/moonshotai.Kimi-K3.1-Pro-GGUF
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of DevQuasar/moonshotai.Kimi-K3.1-Pro-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

