On this page
As of 2026-11-05 Asia/Shanghai, Moonshot’s official Hugging Face repository moonshotai/Kimi-K3.1-Pro is publicly reachable: the model API reports private=false and gated=false. The model card calls Kimi K3.1 Pro an open-weight, natively multimodal agentic model—the first entry in the K3.1 series—with a Mixture-of-Experts design, 2.8T total parameters, 104B activated, and a context length of 1048576 tokens (about 1M). At 2026-11-05 00:00:35 Asia/Shanghai, the official X account @Kimi_Moonshot posted “Releasing the model weights of Kimi K3.1 Pro” and linked this HF repo (status/2118010528123773297). DigIntoAI has not downloaded the ~1.56TB weight tree and has not run the checkpoint locally with vLLM, SGLang, or TokenSpeed. Every figure below follows the Hugging Face model card, the model API, the in-repo LICENSE, the official X post, the ModelScope same-id file list, and Moonshot’s own tech-blog wording—not DigIntoAI measurements.
Related DigIntoAI notes: Kimi Adds K3.1 Pro, Kimi K3.1 Pro API Pricing (October 2026), Kimi Stops Opening K2 to New Users on Oct 31; Full Sunset Planned for Late December, and MiniMax M3.1 Flash Open Weights Are on Hugging Face. The product overview and API/pricing pieces cover hosted Chat / Work / Code / platform routes; what follows stays on open weights on Hugging Face.
What the Hugging Face card and API show
| Field | Value (HF card / API) |
|---|---|
| Repo id | moonshotai/Kimi-K3.1-Pro |
| Visibility | private=false, gated=false |
| Architecture | Mixture-of-Experts (MoE) |
| Total parameters | 2.8T |
| Activated parameters | 104B |
| Context length | 1048576 (~1M) |
| Vision encoder | MoonViT-V2 |
| Pipeline tag | image-text-to-text |
| License (card) | other / license_name=kimi-k3.1-pro (badge: Kimi K3.1 Pro License) |
| Weight files | ~96 model-*-of-000096.safetensors + index |
| Approx. tree size | ~1.56TB (usedStorage ≈ 1,561,355,732,732 bytes when checked) |
createdAt / lastModified | 2026-11-02T04:13:51.000Z = 2026-11-02 12:13:51 Asia/Shanghai |
The HF timestamps are 2026-11-02, not a claim that the shards first appeared on the 2026-11-05 Asia/Shanghai official-X release note. The check day is 2026-11-05 Asia/Shanghai: the repo was already public with those stamps before the X post. Download counters on the API move; DigIntoAI is not treating any snapshot download count as a fixed authority.
Open weights versus the hosted platform API
The same README section that points at the weight recipes also notes the hosted path: on platform.kimi.ai, select model id kimi-k3.1-pro (OpenAI-/Anthropic-compatible API). That hosted route is not the same fact as “weights are open on Hugging Face.” Listing or pricing a platform / OpenRouter id does not, by itself, mean the safetensors tree is downloadable—and the reverse is also true: a public HF repo does not mean DigIntoAI verified a free managed inference endpoint on the hub page.
For product surfaces (Chat, Work, Code, Highspeed, 3D Make, quotas) see the overview Kimi Adds K3.1 Pro. For hosted API list prices see Kimi K3.1 Pro API Pricing (October 2026).
Recommended local inference engines (card copy)
Under Deployment and Availability, the README recommends running Kimi K3.1 Pro on:
- vLLM — recipes at
recipes.vllm.ai/moonshotai/Kimi-K3.1-Pro - SGLang — cookbook at
docs.sglang.io(Moonshotai / Kimi-K3.1-Pro) - TokenSpeed — Lightseek recipes for Kimi K3.1 Pro
Those links are vendor-published runbooks, not DigIntoAI deploy logs. Standing up a ~1.56TB MoE multimodal checkpoint still needs matching hardware, a working runtime, and whatever flags the chosen engine’s recipe documents. DigIntoAI has not verified end-to-end boot, VRAM, throughput, or quality on any of the three stacks.
License: Kimi K3.1 Pro License (not a blank MIT)
The in-repo LICENSE file is titled Kimi K3.1-Pro License (Copyright (c) 2026 Moonshot AI). README badge copy uses Kimi K3.1 Pro License (spacing differs; same license family). Card metadata uses license=other with license_name=kimi-k3.1-pro.
Among the stated conditions (LICENSE full text):
- Permission is granted to use, copy, modify, distribute, and create derivatives, subject to the LICENSE terms.
- If the Licensee (or affiliates) operates a Model as a Service business and aggregate revenue exceeds USD 20 million over any consecutive 12 months, a separate agreement with Moonshot AI is required before commercial use of the Software or derivatives under that MaaS path.
- Large commercial products / services above stated MAU or monthly-revenue thresholds must prominently display “Kimi K3.1-Pro” on the product UI (exact thresholds in the LICENSE).
- Internal use that does not expose the Software’s capabilities to third parties, and use through Moonshot’s official products or certified inference partners, are carved out of some of those commercial display / MaaS clauses—again, as written in the LICENSE, not as a DigIntoAI reinterpretation.
Self-use and internal deploy are not the same compliance story as a high-revenue MaaS offering. The LICENSE text is the source of record.
Official X release note; tech blog still future tense; ModelScope mirror complete (batch 2/2)
At 2026-11-05 00:00:35 Asia/Shanghai, @Kimi_Moonshot posted “Releasing the model weights of Kimi K3.1 Pro” and pointed the primary link at Hugging Face moonshotai/Kimi-K3.1-Pro (X status). That is the official public release note for the weights; it does not rewrite the tech blog’s sentences.
Moonshot’s tech blog at kimi.ai/blog/kimi-k3-1-pro still contains future-tense weight lines such as “The full model weights will be released by Nov 4.” and “Full model weights will be released by Nov 4, 2026.” JSON-LD dateModified remained 2026-10-24 when checked on 2026-11-05. The blog page itself still does not announce “weights are already out” in past tense with an embedded HF link for this repo. The HF tree plus the official X post are the release evidence; the blog copy remains future tense.
Around 2026-11-05 00:02:39 Asia/Shanghai (~two minutes after the X post), ModelScope created a same-id repo moonshotai/Kimi-K3.1-Pro (page open; API Code 200; Id 1627422; ModelScope model page). The same-day upload ran in two SDK batches: batch 1/2 landed at 2026-11-05 03:27:56 Asia/Shanghai (about half the shards then). By 2026-11-05 04:04:25 Asia/Shanghai, the latest commit message reads “Upload folder using ModelScope SDK (batch 2/2)”. The recursive file list now shows 96／96 weight shards—model-00001-of-000096.safetensors through model-00096-of-000096.safetensors—plus model.safetensors.index.json, README／LICENSE／config and related files (~114 paths); StorageSize is 1561355967586 bytes ≈ 1.56TB, aligned with Hugging Face usedStorage (1561355732732). The domestic ModelScope mirror now carries the same full tree that HF hosts. The official release anchor and primary open-weights path named by the X post remain Hugging Face (~1.56TB / 96 safetensors + index; downloads 797 when checked; lastModified still 2026-11-02T04:13:51.000Z; gated=false). DigIntoAI has not downloaded the tree from ModelScope or HF.
What DigIntoAI has not done
- No download of the ~1.56TB safetensors tree
- No local or cluster run with vLLM / SGLang / TokenSpeed
- No re-run of the card’s benchmark tables
- No legal opinion on LICENSE scenarios beyond quoting the published thresholds
- No past-tense rewrite of the tech blog seen
- ModelScope same-id repo finished SDK batch 2/2 (96／96 safetensors + index; ≈1.56TB; aligned with HF)—DigIntoAI still has not downloaded either mirror
Quick facts
- Repo.
moonshotai/Kimi-K3.1-Proon Hugging Face — public, not gated. - Scale. MoE 2.8T total / 104B activated; context 1048576 (~1M).
- Size. ~96 safetensors shards; ~1.56TB.
- Clock. HF
createdAt/lastModified2026-11-02 12:13:51 Asia/Shanghai; checked 2026-11-05. - License. Kimi K3.1 Pro License (LICENSE title Kimi K3.1-Pro License); MaaS revenue threshold USD 20M / 12 months → separate agreement (LICENSE full text).
- Local engines (card). vLLM / SGLang / TokenSpeed.
- Hosted API ≠ weights. Platform id
kimi-k3.1-prois a separate hosted path. - X / Blog / ModelScope. @Kimi_Moonshot “Releasing…” at 2026-11-05 00:00:35 Asia/Shanghai → HF; blog still “by Nov 4”; ModelScope same-id finished batch 2/2 (96／96＋index; ≈1.56TB; domestic full tree now available; HF remains the release anchor).
- Not measured yet. DigIntoAI has not downloaded or run the weights locally.
Bottom line
On the 2026-11-05 Asia/Shanghai check, Moonshot’s official moonshotai/Kimi-K3.1-Pro weights are public on Hugging Face (open-weight MoE, 2.8T/104B, ~1M context, Kimi K3.1 Pro License, ~1.56TB). The HF clock on the repo is 2026-11-02; @Kimi_Moonshot publicly posted “Releasing the model weights of Kimi K3.1 Pro” at 2026-11-05 00:00:35 Asia/Shanghai with the HF link; the tech blog’s weight sentences still read as future tense (“by Nov 4”); ModelScope’s same-id repo finished SDK batch 2/2 at 2026-11-05 04:04:25 Asia/Shanghai (96／96 safetensors＋index; StorageSize ≈1.56TB, aligned with HF)—the domestic mirror now matches the full HF tree, while HF remains the official release anchor. Hosted platform API access and open-weight download remain different facts. DigIntoAI has not downloaded or run the checkpoint—figures above follow the HF model card/API, the official X post, the ModelScope file list, and Moonshot’s published LICENSE and blog copy.