Moonshot AI kept the date. When Kimi K3.1 Pro launched on October 23, the company called it its most capable open-weight model and promised the full weights by November 4. On November 4 the weights arrived: the Hugging Face repository went public during the day, and Moonshot announced the release on X at 16:00 UTC. Anyone can now download the model that, until this week, existed only as an API and an app.

That closes one question and opens a more practical one. The files are real, and they are enormous.

A download measured in terabytes

The checkpoint is about 1.56 terabytes, split across 96 safetensors shards, stored as MXFP4 weights from quantization-aware training. That is roughly comparable to the Kimi K3 release in July, suggesting that the hardware barrier has not materially changed. One third-party hardware guide, VRAM Glass, estimates about 861 GB of VRAM for a 2-bit build and about 1,509 GB at 4-bit, and says running the original weights with vLLM takes eight data center GPUs with 288 GB each. Those figures are estimates, calculated from K3's files because no GGUF build of K3.1 Pro exists yet, and they should be read that way. The download alone is approximately 1.56 TB. Running the model is a separate hardware problem: VRAM Glass estimates roughly 861 GB for a 2-bit build and 1,509 GB at 4-bit, with its vLLM estimate requiring eight 288 GB data-center GPUs for the original weights. These are estimates derived from K3 files, not measurements of a K3.1 Pro deployment.

The practical consequence arrived on the same day. OpenRouter, which routes to hundreds of models, still serves K3.1 Pro only from Moonshot itself, according to a release tracker updated on November 4. Weights being public and third parties hosting the model are two different events. The second one has not happened yet.

The license, read as written

The license was the other missing piece in October, and it is now published. The Kimi K3.1 Pro License grants broad rights to use, copy, modify, fine-tune, distribute and sell, with two conditions aimed at the largest operators. A business that sells model access as a service and takes in more than US$20 million in revenue over any consecutive 12 months must sign a separate agreement with Moonshot before using the model commercially. A product with more than 100 million monthly users, or more than US$20 million in monthly revenue, must display the model's name in its interface. Internal use is exempt.

Two cautions belong next to that summary. Hugging Face tags this license as "other": it is Moonshot's own text, matching the K3 license word for word apart from the model's name, not an OSI-approved license such as Apache 2.0 or MIT. And the US$20 million figure appears twice with different units — revenue in a year for hosting businesses, revenue in a month for the naming rule. A company anywhere near either line should read the text itself, not a summary, this one included.

The scores are still the maker's

Everything on the model card is Moonshot's own measurement, with the harness noted per row. Its headline result is 65.7% on Terminal-Bench 4.0. In Moonshot's reported comparison, Claude Opus 5.5 scored 66.4% and GPT-6 Astra 58.2%. The model card identifies the evaluation harness used for each result; these remain Moonshot-reported figures, not an independent comparison. The card also carries the cyber figures for the released checkpoint, the one trained after cyber-related data was removed: 81.0% on Moonshot's internal exploit benchmark, against 90.5% for the unfiltered checkpoint the company described at launch, plus 74.7 on ExploitBench and 20.7 on ExploitGym.

Independent evaluation of the hosted model exists from October. Independent runs on these released weights do not, yet. That gap matters because a downloaded checkpoint, served on someone else's stack, is the product the open-weight claim is actually about. Until a third party runs it, the strongest honest statement is the narrow one: the weights are out, on time, under terms most developers will never trip over.

Sources

  1. [1] Moonshot AI, Kimi K3.1 Pro model card and Kimi K3.1 Pro License, Hugging Face repository moonshotai/Kimi-K3.1-Pro (official; read first-hand at Gate 1 — repository public and not gated, 96 safetensors shards totaling about 1.56 TB, model card specs and benchmark table, License file text and triggers)Read source
  2. [2] CellCog, "Kimi K3.1 Pro: Open Weights Out, Price and Benchmarks" (updated November 4, 2026; repository first seen public 15:31 UTC, Moonshot X announcement 16:00 UTC, first commit dated November 2, model-card table, OpenRouter still Moonshot-only)Read source
  3. [3] Nowrap, "Kimi K3.1 Pro open weights: download size, license, scores" (rechecked November 4, 2026; no public repository at 04:05 UTC, live at 18:04 UTC, 1.56 TB, license summary, Hugging Face license tag "other")Read source
  4. [4] ExplainX, "Kimi K3.1 Pro: Benchmarks, $2/$10 Price, Open Weights Nov 4" (updated November 4, 2026; 896 experts, MoonViT-V2, license conditions, API pricing)Read source
  5. [5] ArtificialWatch, "Kimi K3.1 Pro" (verified live November 4, 2026, alert 18:16:24 UTC; 96 files, 1.56 TB, not gated, license word-for-word the K3 license)Read source
  6. [6] RuntimeWire, "Moonshot AI releases Kimi K3.1 Pro weights for its 2.8-trillion-parameter model" (published November 4, 2026, 10:36 CT; Moonshot X announcement as primary source)Read source
  7. [7] VRAM Glass, "Kimi K3.1 Pro VRAM Requirements to Run Locally" (third-party estimates, labeled as such in the article; about 861 GB at 2-bit, about 1,509 GB at 4-bit, eight 288 GB data center GPUs under vLLM, no GGUF yet)Read source
  8. [8] ALFW, "Kimi K3.1 Pro Ranks Fourth on Artificial Analysis. Its Weights Aren't Out Yet." (October 23, 2026) — the launch piece that set November 4 as the date to judge the open-weight claimRead source