Kimi K3.1 Pro Open Weights Guide: Use It Without a Server

Moonshot AI has kept its promise: the full weights of Kimi K3.1 Pro, its flagship open model launched on 23 October 2026, are now downloadable from Hugging Face. The release turns one of the strongest coding-agent models of the year into something anyone with serious hardware can inspect, fine-tune and run — but the download itself is about 1.56 TB, and that single number decides almost everything about how you should actually use this model. This guide works through both paths: the hosted route most readers should take, and the self-hosted route for teams that genuinely need the weights.

Quick answer: Kimi K3.1 Pro's open weights are live. Most readers should use the hosted API, not the 1.56 TB download — this guide covers both paths and the hardware bar.

What actually landed on 4 November?

Moonshot launched Kimi K3.1 Pro on 23 October 2026 as its “most capable 3T-class open-weight model to date”, with hosted access from day one and a promise that the full weights would follow by 4 November. Hugging Face lists the repository moonshotai/Kimi-K3.1-Pro as created on 2 November, and independent watchers found it public during 4 November, the promised date, with Moonshot announcing the release on X the same day. Our earlier launch coverage is here: https://openaimaster.ai/kimi-k3-1-pro-launch/

https://x.com/Kimi_Moonshot/status/2118010528123773297

The model card on Hugging Face describes Kimi K3.1 Pro as an “open-weight, natively multimodal agentic model”. Its headline specifications, all from Moonshot’s own card, are:

Specification	Kimi K3.1 Pro
Architecture	Mixture-of-Experts (MoE)
Total / active parameters	2.8 trillion / 104 billion per token
Experts	896, with 16 selected per token (plus 2 shared)
Context window	1,048,576 tokens (1 million)
Input	Text and images
Weights	MXFP4 weights, MXFP8 activations (quantization-aware training)
Vision encoder	MoonViT-V2, 401 million parameters

The download is reported at about 1.56 TB spread across 96 safetensors shards. That is a server-class download, not a laptop one — and it is the first filter in the decision below. Note the “3T-class” label is Moonshot’s own class name, not the parameter count: the card says 2.8 trillion total parameters.

Should you download the weights at all?

For most readers, no — and that is not a criticism of the release. Downloading makes sense only if you need something the hosted model cannot give you: running the model inside your own infrastructure, fine-tuning it on private data, or inspecting the actual files rather than trusting an API. If you just want Kimi K3.1 Pro’s answers, the hosted model is the same model, without the download, the GPUs or the operations work.

Your situation	Best path
You want to try K3.1 Pro, build with it, or use it for coding work	Hosted: Kimi website, app, Kimi Code, or API
You need it inside your own infrastructure or data boundary	Self-hosted weights, if you clear the hardware bar below
You want to fine-tune it on your own data	Self-hosted weights (check the licence first)
You run a laptop or a single desktop GPU	Hosted — the weights are not built for that hardware

Path 1: Use Kimi K3.1 Pro without downloading anything

This is the route we recommend starting with, whatever your long-term plan. Kimi K3.1 Pro is live on Kimi’s website and apps, in Kimi Work, in Kimi Code (select it with the /model command), and through the Kimi API as kimi-k3.1-pro, with OpenAI- and Anthropic-compatible interfaces, according to Moonshot’s model card. A faster Highspeed variant is also available.

API pricing, from Moonshot’s pricing page as reported in launch coverage, is $2 per million input tokens and $10 per million output tokens, with cached input at $0.20 per million; Highspeed costs $4 and $20. That is a third less than Kimi K3’s $3 and $15, and it is the number to use for the cost comparison in the next section.

Two usage details from the model card are worth knowing before you build anything. First, thinking effort is set with a reasoning_effort field that accepts “low”, “high” or “max”, and defaults to “max” — so the model’s most expensive reasoning setting is the default, not an upgrade. Second, for multi-turn conversations and tool calls, the card requires passing back the complete assistant message as returned, including reasoning_content and tool_calls, not just the visible content; dropping the reasoning content breaks the conversation state the model expects.

Path 2: Run the open weights yourself, step by step

If you cleared the decision table above and self-hosting is genuinely what you need, here is the honest sequence.

Step 1: Read the licence before you download.

Kimi K3.1 Pro is released under its own Kimi K3.1 Pro License, not under MIT or Apache 2.0 — Hugging Face tags the licence as “other”. Licence summaries report broad rights to use, copy, modify, fine-tune, distribute and sell, with two conditions that matter at scale: a company running a model-as-a-service business with more than $20 million in revenue over 12 months must sign a separate agreement with Moonshot before commercial use, and a product with more than 100 million monthly users, or more than $20 million in monthly revenue, must display “Kimi K3.1-Pro” in its interface. Almost every individual and small team falls below both thresholds, but read the licence text itself if you are near either one.

Step 2: Budget storage honestly.

The weights are about 1.56 TB. That is the download alone — you also need working room for the download process, the inference engine and its caches, so plan storage well above 1.56 TB before you start.

Step 3: Pick an inference engine from Moonshot’s own list.

The model card recommends vLLM, SGLang or TokenSpeed, each with a recipe or cookbook linked from the card. Do not improvise a serving stack for a 2.8-trillion-parameter model; start from one of these three.

Step 4: Clear the hardware bar, on paper first.

Reported vLLM recipes run the model across eight GPUs — two Nvidia GB300 trays, one B300 node, or one AMD MI355X node, according to MadRobot’s write-up of the release, with one report putting the GPU-memory requirement around 2.3 TB. Treat those figures as reported, not as our own test: we have not run the weights ourselves, and your mileage will depend on the engine, the quantization path and the context lengths you actually serve. If you cannot describe your eight-GPU-class setup before downloading, you are not ready for this path.

Step 5: Set reasoning effort deliberately.

Because “max” is the default, a self-hosted deployment left on defaults runs at maximum reasoning effort on every request. For bulk or latency-sensitive work, test “high” and “low” explicitly and measure what each costs you in tokens and time before committing to a default.

What the benchmarks do and don’t prove

Moonshot’s own model card puts Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 and 58.2% for OpenAI’s GPT-6 Astra. Those are Moonshot’s numbers, evaluated with the Claude Code harness, and the card itself notes where comparison scores come from model blog posts or official leaderboards rather than a single shared run. Independent evaluator Artificial Analysis ranks K3.1 Pro fourth on its Intelligence Index at 55.1. The honest reading: the model is credibly near the frontier on coding-agent work, but until independent labs rerun these tests on the downloadable files, the headline scores remain vendor-reported, not independently reproduced on the weights themselves.

Our take: the 1.56 TB test

Every open-weights release invites the same reflex: download first, decide later. Kimi K3.1 Pro punishes that reflex more than most, because the download alone is a multi-day, multi-terabyte commitment. Our test is five questions, and a “no” to any of them sends you back to the hosted API: Do you need the model inside your own infrastructure? Do you have eight-GPU-class hardware, or a budget to rent it? Have you read the licence thresholds? Can you name the inference engine you will use? And can you beat $2/$10 API pricing once hardware, power and operations time are counted? A worked example makes the last question concrete: a team sending 20 million input tokens and 5 million output tokens a month would pay $40 plus $50 — $90 — through the API at published prices. Self-hosting has to beat $90 a month after hardware and engineering time, which, for most teams, it will not. That arithmetic, not the download button, is the real decision — and it is why the weights matter most as a research and infrastructure story, while the API remains the practical way to use the model. Desk-researched, not hands-on tested: we verified the model card and the release reports, but we did not run the weights ourselves.

Frequently Asked Questions

Are Kimi K3.1 Pro’s weights really free to download?

Yes, the repository is public with no access request, according to multiple release reports. “Free to download” is not the same as “free of conditions”: the custom Kimi K3.1 Pro License carries the revenue and user-count thresholds described above, so commercial users near those thresholds should read the licence text itself.

Can I run Kimi K3.1 Pro on my laptop?

No. At about 1.56 TB of weights and a reported eight-GPU-class serving setup, this is a data-centre model. Use the hosted website, app or API on a laptop instead — that is the same model, without the hardware.

Do I need the weights to use Kimi K3.1 Pro at all?

No. The hosted model has been available since the 23 October 2026 launch on Kimi’s website, apps, Kimi Work, Kimi Code and the API. The weights release adds self-hosting, fine-tuning and inspection — it does not unlock a different model.

What does the thinking-mode setting do?

Kimi K3.1 Pro’s thinking effort is controlled by the reasoning_effort field (“low”, “high”, “max”, default “max”). Higher effort means more reasoning before an answer, which usually means more tokens and more time. For multi-turn and tool use, you must pass back the full assistant message, including its reasoning content, or the model loses conversation state.

How is this different from your 23 October launch article?

The launch article covered the hosted release: the model, its price and its first benchmarks. This guide covers what changed on 4 November — the downloadable weights, the licence, and the practical choice between the API and self-hosting. Read the launch piece first for the full launch picture: https://openaimaster.ai/kimi-k3-1-pro-launch/

Sources and methodology

Sources, each linked in the article or below (all open in a new tab): Moonshot AI’s Kimi K3.1 Pro model card on Hugging Face (https://huggingface.co/moonshotai/Kimi-K3.1-Pro) for specifications, engines, reasoning effort and licence name; MadRobot’s 23 October launch article updated 4 November 2026 (https://madrobot.blog/2026/10/23/kimi-k3-1-pro-moonshot-open-weight-nearly-matches-claude-opus-5-5-coding/) for the weights timeline, API pricing and hardware recipes; Nowrap’s 4 November recheck (https://nowrap.ai/news/moonshot-kimi-k3-1-pro-launch) for the repository’s public timing and licence summary; and Moonshot’s X announcement of the weights release (https://x.com/Kimi_Moonshot/status/2118010528123773297). Related reading on this site: our Kimi K3.1 Pro launch article (https://openaimaster.ai/kimi-k3-1-pro-launch/) and our MiniMax M3.1 guide (https://openaimaster.ai/minimax-m3-1-flash-guide/) for a parallel weights-era walkthrough. Method: every factual claim was verified fresh on 5 November 2026 against the primary model card first and independent secondary reporting second; vendor benchmark figures are labelled as Moonshot’s own, licence thresholds are reported from licence summaries rather than quoted from the full licence text, and hardware figures are reported from release coverage, not from our own testing — this article is desk-researched, not hands-on tested.

Govind Dheda

Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.