Kimi K3.1 Pro open-weight AI model release illustration showing a neural network releasing weight files to servers worldwide

Moonshot Releases Kimi K3.1 Pro Weights: The Biggest Open-Weight AI Model You Can Download

ByChris Wade

Key takeaway: Moonshot AI released the weights of Kimi K3.1 Pro on Hugging Face on November 4, 2026 — a 2.8-trillion-parameter open-weight model with a 1-million-token context window, the largest publicly downloadable model of its kind. Anyone can download it, but the 1.56 TB download and a revenue-gated licence mean most people will still use it hosted.

What changed on November 4

Moonshot AI posted “Releasing the model weights of Kimi K3.1 Pro” on November 4, 2026, at 16:00 UTC, linking to the public Hugging Face repository moonshotai/Kimi-K3.1-Pro. The release met the deadline Moonshot set when it launched the model on October 23, 2026: the weights would be out “by Nov 4.” The repository is not gated — no access request is needed — and it holds 96 weight files totalling about 1.56 TB. The model is also available through Moonshot’s own API as kimi-k3.1-pro, and hosted versions appeared on Hugging Face, Ollama Cloud, Nebius, and Verda on day zero.

Kimi K3.1 Pro is the first model in Moonshot’s new K3.1 series, post-trained from the Kimi K3 base model with large-scale joint reinforcement learning across hundreds of tasks. It is built for agentic work — coding, 3D tasks, financial research, and other jobs where the model takes many steps on its own. Our report on Tencent’s Hy4 launch covers the other big open-weight release of the same week.

What’s inside the model

Kimi K3.1 Pro is a mixture-of-experts (MoE) model: 2.8 trillion parameters in total, but only 104 billion — about 3.7% — are active for each token. It has 896 experts, with 16 selected per token, and a context window of 1,048,576 tokens, the same 1-million-token window as Kimi K3. It takes text and image input natively, using a 401-million-parameter MoonViT-V2 vision encoder. The weights use MXFP4 precision with MXFP8 activations, trained with quantization awareness, and 24 of its 93 layers replace full attention with Gated DSA (DeepSeek Sparse Attention) to speed up long-context work.

Model Total parameters Active per token Context window Weights
Kimi K3.1 Pro (Moonshot) 2.8 trillion 104 billion 1,048,576 tokens Public, Nov 4, 2026
Kimi K3 (Moonshot) 2.8 trillion Not disclosed 1,048,576 tokens Public since July 2026
Hy4 (Tencent) 770 billion 49 billion Not disclosed Promised, not yet released
GPT-6 Astra (OpenAI) Not disclosed Not disclosed Not disclosed Closed, API only

The table’s contrast matters: Kimi K3.1 Pro is the only model here whose full weights you can actually download today. Hy4’s weights are still promised, and GPT-6 Astra is closed. That is what makes this release the week’s most practical open-weight news, even though the model is not the largest by active parameters.

How does it perform?

All benchmark figures below are Moonshot’s own, reported on the model card — no independent reproduction existed at publication time, so treat them as vendor claims. Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. Artificial Analysis rates it at 55 points, fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56). In Moonshot’s own benchmark table, Kimi K3.1 Pro scores above Claude Opus 5.5 on 16 of the 37 tests where both have scores, with its largest lead at 5.4 points — though Opus 5.5 leads by more than 20 points on three rows. Other vendor-reported scores include 94.2 on GPQA-Diamond, 92.1 on BrowseComp, and 79.6 on OSWorld V2.1.

What does it cost to run?

There are two costs: the hosted price and the self-hosted hardware. Through Moonshot’s API, Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens. A faster variant, Kimi K3.1 Pro Highspeed, runs at 2–3 times the speed — but its weights were not released, so it is API-only. Self-hosting is a different scale: the download alone is 1.56 TB, and the simplest vLLM setup needs eight data-centre GPUs. Moonshot lists vLLM, SGLang, and TokenSpeed as recommended inference engines, with deployment recipes for vLLM and SGLang published alongside the weights. For most readers, the honest answer is hosted access: the weights matter most to researchers, labs, and companies that need to run the model inside their own infrastructure.

What you need before you download

Four prerequisites come before the download button. First, storage and bandwidth for 1.56 TB across 96 files. Second, GPU hardware — eight data-centre GPUs for the reference vLLM setup, or a smaller quantised build from the community copies already appearing on Hugging Face. Third, familiarity with vLLM or SGLang, since Moonshot ships deployment recipes, not a one-click installer. Fourth, and most overlooked, the licence: the weights ship under Moonshot’s own Kimi K3.1-Pro License, not a standard open-source licence. It is MIT-style for most users, but a model-as-a-service business whose revenue with affiliates exceeds $20 million over any 12 months must sign a separate agreement with Moonshot before commercial use, and products with more than 100 million monthly active users or $20 million in monthly revenue must display attribution. Check the licence on the model card before building a product on it.

Knowns, unknowns, and what to watch

Known: the weights are public, ungated, and complete for Kimi K3.1 Pro (Highspeed excluded); the size, context, licence terms, and API price are all confirmed on the model card and Moonshot’s pricing page. Unknown: independent benchmark results — no third-party reproduction had landed when the weights went live, and Artificial Analysis’s score is an index, not a rerun. Also unconfirmed: a report that the model trained on GPT-6 Astra’s outputs. Moonshot has not confirmed it, and it should be treated as an unverified claim, not a fact. What to watch: whether independent labs reproduce the Terminal-Bench lead, how quickly community quantisations shrink the 1.56 TB footprint, and whether Tencent follows through on Hy4’s promised weights — if it does, two 2-trillion-plus-class open-weight models will be downloadable within weeks of each other.

Analysis: the biggest download is not the biggest deal — the licence is

Analysis: Kimi K3.1 Pro’s headline is its size — the largest publicly downloadable open-weight model — but the durable story is the licence. A 2.8-trillion-parameter model anyone can download is a genuine research win, and the 1-million-token context plus strong agent benchmarks make it a real option for coding and long-document work. Yet the $20 million revenue gate and the attribution rule mean the biggest commercial hosts cannot simply adopt it for free, and the 1.56 TB, eight-GPU reality means almost nobody will run the full model at home. Moonshot has opened the weights while keeping a hand on the commercial value — which is a reasonable trade, but it is not the same as a fully open release. Judge it by what you can actually do with it: researchers and self-hosters gain a frontier-class model; everyone else gains a cheaper hosted API.

Frequently asked questions

Is Kimi K3.1 Pro really free to download?

Yes, the weights are public and ungated on Hugging Face at moonshotai/Kimi-K3.1-Pro — no access request needed. But “free to download” is not “free for every use”: the Kimi K3.1-Pro License requires model-as-a-service businesses earning over $20 million in 12 months to sign a separate agreement with Moonshot before commercial use.

How big is the Kimi K3.1 Pro download?

About 1.56 TB, spread over 96 safetensors weight files. You also need serious hardware to run it: the simplest vLLM setup calls for eight data-centre GPUs. Most users will find hosted access through Moonshot’s API or a day-zero host like Nebius or Ollama Cloud far more practical.

How does Kimi K3.1 Pro compare to Claude Opus 5.5?

On Moonshot’s own benchmark table, Kimi K3.1 Pro beats Claude Opus 5.5 on 16 of 37 shared tests, but Opus 5.5 leads by more than 20 points on three rows and scores higher on Terminal-Bench 4.0 (66.4% vs 65.7%) and on Artificial Analysis’s index (58 vs 55). The two are close on agentic coding, and all figures are vendor-reported until independent labs reproduce them.

What is the Kimi K3.1 Pro API price?

Moonshot charges $2 per million input tokens and $10 per million output tokens for kimi-k3.1-pro. A faster variant, Kimi K3.1 Pro Highspeed, runs at 2-3 times the speed but is API-only — its weights were not released.

What is the difference between Kimi K3.1 Pro and Kimi K3?

Kimi K3.1 Pro is post-trained from the Kimi K3 base model with large-scale joint reinforcement learning, and it adds Gated DSA sparse attention in 24 of 93 layers for faster long-context work. Both are 2.8-trillion-parameter models with a 1-million-token context, but K3.1 Pro is the newer, more capable agentic model in the K3.1 series.

Sources


Similar Posts