What GPU do you need to run Kimi K3.1 Pro?
Confirmed setup not yet documented. Official weights are a 1,561.29 GB download, not the memory a machine needs. Community GGUFs now exist, including a preliminary Q4, but this site still estimates neither fit nor speed. A rough planning budget is around 2 TB of aggregate memory—approximately the capacity of four 512 GB Mac Studios. Actual requirements and support for running across those Macs remain unconfirmed.
Parameters2780.11B104B active (mixture-of-experts)
Context window1M tokens
2780.11B parameters, 104B active (mixture-of-experts) (Hugging Face, accessed 2026-11-09). No Q4 GGUF is sourced, so this page does not estimate a fit or a speed. The weight download is 1,561,293,160,568 bytes (1,561.29 GB) across 96 safetensors shards (source). That is the size of the files, not the memory a runtime keeps resident. The whole repo, including the index and tokenizer, is 1,561,355,732,732 bytes. Commit d9e99b0792a5e93b78ea9bea62a68167e36ddd9e. No context estimate: Kimi K3.1 Pro's KDA and Gated DSA attention has no verified llama.cpp KV formula yet (details).
104B active is the model card's count of parameters used for one token. It is not the memory the checkpoint occupies, and the weight download is not that memory either. Server recipes and community uploads are not a confirmed setup for a PC or Mac on this site. Community GGUFs now exist, but their file sizes are downloads, and this site does not use them for a fit or a speed. RAM offload, an in-memory cluster, SSD streaming, and a prefill/decode split are different ways to place a model. None of them is confirmed here, and no speed or rental price is listed. A rough planning budget is around 2 TB of aggregate memory—approximately the capacity of four 512 GB Mac Studios. Actual requirements and support for running across those Macs remain unconfirmed.
What we checked
Checked Nov 9, 2026. Official is the maker or the weight files. A vendor claim is that project's own documentation. A community claim is someone else's upload. Our check is a file or search we read. Unknown means we have no source, so it is not estimated. None of these rows is a speed we measured.
| Kind | What | Detail | Source |
|---|---|---|---|
| Official | Model card | Moonshot's card calls it an open-weight multimodal mixture-of-experts model: 2.8T total parameters, 104B activated per token, 93 layers (1 dense; 69 KDA and 24 Gated DSA), 896 experts with 16 selected per token plus 2 shared, a 1,048,576-token context, and text and image input through MoonViT-V2 (401M parameters). The card's quantization line is MXFP4 weights and MXFP8 activations, from quantization-aware training. Hugging Face's safetensors total is 2,780,110,363,392 parameters (2,780.11B on this page); the card rounds that to 2.8T. The 104B figure is not resident memory. | Model card, checked Nov 9, 2026 |
| Official | Weight files | 96 safetensors shards sum to 1,561,293,160,568 bytes (1,561.29 GB). The repo's used storage, including the index and tokenizer, is 1,561,355,732,732 bytes (1,561.36 GB). The repo was created 2026-11-02 and last modified 2026-11-05 (commit d9e99b0792a5e93b78ea9bea62a68167e36ddd9e, still the tip on 2026-11-09). Both sizes are downloads. They are not GPU memory, unified memory, or a GGUF. | Hugging Face API, checked Nov 9, 2026 |
| Official | Which weights are MXFP4 | config.json sets text_config.quantization_config (there is no top-level quantization_config) to compressed-tensors mxfp4-pack-quantized, 4-bit, group size 32. Its ignore list leaves attention, shared experts, the dense MLP, the router, lm_head, the vision tower, and the multimodal projector out of that group. kv_cache_scheme is null, so the file does not state a KV-cache format. The card's MXFP8-activations line describes the training quant, not a cache size this site can turn into a context estimate. | config.json, checked Nov 9, 2026 |
| Official | License | The weights use the Kimi K3.1-Pro License (license_name kimi-k3.1-pro). It allows use, modification, and redistribution if the copyright notice stays with the files. If you and your affiliates take more than $20 million over any 12 months from offering inference or fine-tuning as a service, you need a separate Moonshot agreement before using the weights commercially. A product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must show the name Kimi K3.1-Pro. Internal use, and use through Moonshot's own products or its certified partners, is outside those two conditions. This is a summary of the license, not legal advice. | LICENSE, checked Nov 9, 2026 |
| Official | Hosted API | Moonshot's API quickstart names the hosted model id kimi-k3.1-pro and a 1M-token context. That is Moonshot's service, not a local setup. This page has no API price. | API quickstart, checked Nov 9, 2026 |
| Official | Tech blog | Moonshot's tech blog introduces Kimi K3.1 Pro and a high-speed API variant, and says the full weights would be released by Nov 4, 2026. The Hugging Face repo was created on Nov 2, 2026. The blog does not name a PC, a Mac, or a measured local speed. | Moonshot tech blog, checked Nov 9, 2026 |
| Vendor claim | vLLM recipe | The model card links this recipe. It says the simplest NVIDIA setup it documents is tensor parallel 8, on two GB300 trays or one B300 node, and the same eight-GPU layout on one MI355X node, for the MXFP4 checkpoint. It says that was verified on NVIDIA GB300 NVL4. The page labels the variant MXFP4 1874 GB and says the routed-expert weights are about 1.56 TB on disk, with attention, shared experts, and the vision tower left in BF16. It also describes prefill/decode disaggregation and a host-memory KV tier as extra server options. Those are vLLM's statements. We have not measured resident memory or speed, and 1874 GB is not a fit in our calculator. | vLLM recipe, checked Nov 9, 2026 |
| Vendor claim | SGLang | SGLang's November 4, 2026 day-0 post says every recipe uses a preview image and that support has not shipped in a stock SGLang release. The post says they ran their own AgentX test of the MXFP4 checkpoint on an NVIDIA GB300 NVL72, including prefill/decode splits, and it names other accelerator families. The model card links the cookbook for launch commands. The speeds in that post are SGLang's. We did not reproduce them and do not list them. | SGLang day-0 post, checked Nov 9, 2026 |
| Vendor claim | TokenSpeed | The model card links TokenSpeed's recipe. That page says the NVIDIA launch was validated on two GB300 nodes with 4 GPUs each, and that the context length is the checkpoint's 1,048,576 tokens. The eight-GPU MI350X or MI355X command still sets --max-model-len 65536, not the full 1,048,576. The DSpark and EAGLE draft-model flags on that page are optional speculative-decoding recipes, not a measured plain-decoding speed. That is TokenSpeed's documentation. We have not run it. | TokenSpeed recipe, checked Nov 9, 2026 |
| Vendor claim | Ollama cloud | Ollama's library page lists kimi-k3.1-pro:cloud. Its examples call ollama.com. That is a hosted model, not a file that runs on your machine. | Ollama library, checked Nov 9, 2026 |
| Our check | Community GGUFs | A Hugging Face search filtered to GGUF now returns several Kimi K3.1 Pro repos. Sizes below are summed from each repo's file listing and are downloads, not GPU or unified memory. DevQuasar/moonshotai.Kimi-K3.1-Pro-GGUF (card says VERY EXPERIMENTAL): MXFP4_MOE, 33 files, 1,508,897,280,480 bytes (1,508.897 GB); Q2_K, 33 files, 1,009,315,027,936 bytes (1,009.315 GB); and a BF16 mmproj of 904,448,928 bytes (0.904 GB). danielhanchen/Kimi-K3.1-Pro-GGUF-Preliminary (card title: Kimi K3.1 Pro Preliminary quants): UD-Q4_K_XL, 34 files, 1,509,020,584,672 bytes (1,509.021 GB); UD-Q8_K_XL, 34 files, 1,561,525,956,960 bytes (1,561.526 GB). Because that Q4 set is labeled Preliminary, this page does not use it as the Q4 file for the calculator. KarLexAI/Kimi-K3.1-Pro-GGUF is text only (the card says the vision encoder is not included): KLX-IQ2_XXS, 17 files, 709,876,827,040 bytes (709.877 GB); KLX-IQ1_M, 15 files, 647,869,832,960 bytes (647.870 GB); KLX-IQ1_S, 14 files, 592,677,343,328 bytes (592.677 GB). That card publishes quality and speed claims for an 8×H100 node; those are the uploader's, and we have not run them. midagedev/Kimi-K3.1-Pro-GGUF: IQ1_S, 31 files, 580,345,966,528 bytes (580.346 GB); the card says only bloomery loads it, not mainline llama.cpp, and its parameter counts do not match Moonshot's 2.8T total and 104B active. unsloth/Kimi-K3.1-Pro is still a safetensors copy of the official 96 shards (1,561,293,160,568 bytes), not a GGUF. bartowski has no Kimi K3.1 Pro repo, and mlx-community has no Kimi K3.1 Pro repo. None of these files is a confirmed setup on this site. | Hugging Face GGUF search, checked Nov 9, 2026 |
| Our check | llama.cpp | PR #31779, "model: add Kimi-K3.1-Pro text model", merged on 2026-11-06 and added text support under the architecture name kimi-k31. Vision support, PR #31980 ("mtmd: add kimi K3.1 vision"), merged on 2026-11-10. Both are in tagged release v1.0.0 (2026-11-13), whose release notes list Kimi-K3.1-Pro (#31779) among the new text architectures and Kimi-K3.1 vision (#31980). We have not checked whether Ollama or LM Studio can load it. A merged loader is not a confirmed setup or a speed on this site. | llama.cpp v1.0.0 release, checked Nov 13, 2026 |
| Our check | mlx-lm | mlx-lm main at commit 56230fa02b8ff8330daf26dc95c28480423ed5fe (2026-11-05) is still the tip of main on 2026-11-09. Its models folder has kimi_k3.py, kimi_k25.py, kimi_linear.py, and kimi_vl.py, and no kimi_k31.py. Upstream mlx-lm is not a working loader for this checkpoint. | mlx-lm models folder, checked Nov 9, 2026 |
| Community claim | inferencerlabs MLX upload | 115 safetensors files sum to 1,561,243,578,065 bytes (1,561.244 GB); the repo was last modified 2026-11-05, after this page's earlier 112-file snapshot. The card says the build repacks the existing quantized weights and offloads the experts to storage, names a 512 GB M3 Ultra, and now publishes a speed figure. Storage offload is not the same as holding the model in GPU memory or in a Mac's unified memory. That speed figure is the uploader's, not a measurement on this site, and we have not run it. | inferencerlabs card, checked Nov 9, 2026 |
| Community claim | AinsleyMarkov 2-bit upload | 95 safetensors files sum to 816,943,261,367 bytes (816.94 GB). The card says routed experts are stored at 2 bits and that the file is meant for the author's tensor-parallel MLX engine, which is not published. It also says mlx-lm has no kimi_k31 model. The 816.94 GB figure is a download size. It is not evidence that a Mac, or any machine we track, can run it. | AinsleyMarkov card, checked Nov 9, 2026 |
| Community claim | pipenetwork REAP80 MLX upload | pipenetwork published REAP-pruned MLX builds that keep 179 of 896 routed experts per MoE layer and ship a custom kimi_k31.py loader; the card says the kimi_k31 architecture is not upstream in mlx-lm. The q6 build, the uploader's recommended tier, has 95 safetensors totaling 335,960,941,648 bytes (335.961 GB). The same family has a q4 build of 321,897,117,664 bytes (321.897 GB) and a q8 build of 350,024,765,392 bytes (350.025 GB). The card publishes quality and speed claims for a 512 GB M3 Ultra. Those claims are the uploader's. We have not run these files, and a pruned build with a custom loader is not a confirmed setup on this site. | pipenetwork card, checked Nov 9, 2026 |
| Community claim | dphn NF12 upload | 96 safetensors files sum to 1,532,965,277,463 bytes (1,532.965 GB). The card describes an NF12 compression of the checkpoint's BF16 tensors, aimed at an 8-GPU B300-class server with the author's runtime. It is not a GGUF or an mlx-lm build, and we have not run it. | dphn card, checked Nov 9, 2026 |
| Unknown | Not documented | Unknown, and not estimated here: how much memory a runtime keeps resident, KV-cache bytes per token, any measured speed, a tested PC or Mac in our catalog, RAM offload of the weights, an in-memory cluster of machines we list, and a current rental price. Prefill/decode splits appear only as options inside the server recipes above. They are not a confirmed setup for one machine. | Model card does not give this, checked Nov 9, 2026 |
Other places to get it
Checked Nov 14, 2026. We list only copies we opened and checked. Check file hashes for copies from other uploaders.
- ModelScope: Lab's own account
- Ollama: Cloud only (runs on Ollama's servers)
ollama run kimi-k3.1-pro:cloud