Kimi K3.1 Pro’s weights are out. What it takes to run, and how to use it now

What changed on November 4

On November 4, 2026, Moonshot AI posted “Releasing the model weights of Kimi K3.1 Pro,” on the deadline it set at launch. They’re on Hugging Face at moonshotai/Kimi-K3.1-Pro with no access gate, 1.56 TB in 96 files by Hugging Face’s record. Artificial Analysis’s model page now reads: “The model weights are publicly available and can be downloaded for self-hosting.”

What that means depends on how you use the model:

• In Kimi’s apps, Kimi Code or the API: nothing changes, and the prices stay as they were.
• On your own hardware: it’s possible now, if you have eight data-center GPUs. A laptop won’t do.
• In a product you sell: you can build on it. If you resell inference or you’re very large, read the license’s two conditions below.

What Moonshot released

Moonshot announced K3.1 Pro on October 23 as “the first model in our new K3.1 series and our most capable 3T-class open-weight model to date,” with a context window of one million tokens. It’s a post-trained version of Kimi K3, Moonshot’s previous large model, made with what the company calls “substantially larger post-training compute.” The model card gives its size as 2.8 trillion parameters in total, 104 billion of them active for each token, with text and image input.

It’s also on Kimi.ai and the Kimi mobile apps, in the Kimi Work desktop app, in Kimi Code, Moonshot’s coding agent, and on the Kimi API as kimi-k3.1-pro. A second version, Kimi K3.1 Pro Highspeed (kimi-k3.1-pro-highspeed), runs at “2-3× the speed” of the first. The weights cover K3.1 Pro: Highspeed has no repository of its own, and the card doesn’t mention it.

What it takes to run

A data center, or a rented piece of one. Moonshot’s card gives no memory or GPU figures. It recommends three engines, vLLM, SGLang and TokenSpeed, and leaves the hardware to their recipes:

What	Figure	Source
Download	1.56 TB	Hugging Face
Simplest NVIDIA setup	8 GPUs, as one B300 node or two GB300 trays	vLLM
AMD setup	One MI355X node	vLLM
vLLM version	0.34.0 or later	vLLM
An 8-GPU B300 system	2.1 TB of GPU memory, about 14 kW	NVIDIA
Renting 8 B300s	$71.92 an hour at $8.99 a GPU, about $52,500 for a 730-hour month	Runpod

vLLM calls the eight-GPU layout “the simplest working setup on NVIDIA” and adds: “Throughput was not tuned.” The checkpoint is already quantized, with 4-bit weights for its routed experts, and Moonshot offers no smaller version. vLLM and SGLang serve it over an OpenAI-compatible endpoint, which is how a coding agent connects to it.

vLLM’s command includes --trust-remote-code, which runs Python from the repository on your machines, so download from Moonshot’s account rather than a copy. Within an hour of Moonshot’s post, Hugging Face listed re-uploads from other accounts, and one repository using the model’s name dates from October 23, before the weights existed.

There’s no cheap route yet. We found no GGUF build of K3.1 Pro on Hugging Face on November 4. For Kimi K3, a model of the same size, Unsloth’s guide recommends a 594 GB build that “will require at least 610GB RAM.” A K3.1 Pro build, if one comes, is unlikely to be much smaller, though nobody has published one. For a coding model that fits on a laptop, see Flint or our guide to running a coding agent on a local model.

Self-hosting isn’t how to save money at small scale. By our arithmetic, a $52,500 month of rented GPUs buys about 13.1 billion input tokens and 2.6 billion output tokens on Kimi’s API, at five input tokens for each output token and with no caching, before you count storage, networking or staff. The case for it is keeping your data in your own hands, or steady traffic at that volume. Other hosts haven’t undercut Moonshot so far: OpenRouter lists Moonshot, Together and Phala at the same $2 and $10, and Together’s own page still carried a “Launching soon” banner on November 4.

What the license allows

The Kimi K3.1-Pro License is, apart from the model’s name, word for word the Kimi K3 License. It isn’t a standard open-source license, but for most people it works like a permissive one: you may use, copy, change, fine-tune and sell the model, as long as you keep its copyright notice. Two conditions apply at scale:

1. Hosting it for others. A business giving third parties “language model inference or fine-tuning” with “meaningful control over the inputs, parameters, or training data,” whose group revenue passes $20 million over any 12 months, “must enter into a separate agreement with Moonshot AI” before commercial use. Features embedded in an end-user product don’t count as that kind of service.
2. Credit at very large scale. A product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must prominently display “Kimi K3.1-Pro” in its interface.

Neither applies to internal use, defined as “any use that does not make the Software, its outputs, or its underlying capabilities available to third parties,” or to use through Moonshot’s own products or “certified inference partners.” The model card doesn’t name those partners.

How good it is

Moonshot ran most of its coding tests in Claude Code’s harness, with reasoning effort at its maximum. Here are its own figures against Claude Opus 5.5 and GPT-6 Astra, from the model card:

Benchmark	Kimi K3.1 Pro	Claude Opus 5.5	GPT-6 Astra
Terminal-Bench 4.0	65.7%	66.4%	58.2%
SWE Marathon	51.7%	46.3%	42.5%
Kimi Code Bench 3.0 Long-horizon	61.2%	58.5%	66.5%
Terminal-Bench Science 0.1	38.1%	63.3%	68.1%

Kimi Code Bench is Moonshot’s internal test, and on it GPT-6 Astra still leads.

Two outside testers have run it. Artificial Analysis ranked it fourth at launch, at 55 on version 4.3 of its Intelligence Index, and measured 67% on Terminal-Bench 4.0, “the highest score of any model” on that test at the time. On October 27 it replaced that index with version 5, built mostly on private tests. On today’s leaderboard, version 5.0.1, K3.1 Pro scores 46, against 59 for Opus 5.5 and 60 for Gemini 4 Carbon at their top efforts. That’s still the highest of any open-weights model there: DeepSeek V4.1 Flash 1017 follows at 33 and GLM-5.3 at 32. An index task costs $3.68 at API prices.

Vals AI gives it 62.52% on its Vals Index, ninth of 32 and what its update calls “the top open-weights model,” and 63.64% on Terminal-Bench 4.0, third of 59. It scored below Kimi K3 on Vals’s legal research and tax tests, so the gains aren’t across the board.

It’s also slow on Moonshot’s own API. Artificial Analysis measures 32 tokens per second, 146 seconds to the first answer token and 37.1 minutes for an index task.

What Moonshot took out

For the open-weight release, Moonshot says it removed cyber-related data from training. On xdev-bench-hard, its internal vulnerability-research test, K3.1 Pro solves 81.0% of problems, against 90.5% for the same training without the cuts. On the card’s ExploitBench it scores 74.7, against 100 for GPT-6 Astra and 91.0 for Opus 5.5, with scores from each company’s own reports. Moonshot says K3.1 Pro’s cyber capability “remains behind that of frontier proprietary models that have been publicly available for at least three months,” and that “we aim not to release open weights that would extend the absolute frontier of cyber capability.”

What it costs

On the API, per million tokens:

Model	Input	Output
Kimi K3.1 Pro	$2	$10
Kimi K3.1 Pro Highspeed	$4	$20
Kimi K3	$3	$15
Claude Sonnet 5.5	$2	$10
Claude Opus 5.5	$4	$20
GPT-6.1 Sol	$2	$10

Kimi’s rates are from its K3.1 Pro pricing and API pricing pages, and cached input costs $0.20 on both K3.1 models. The others are from Anthropic and OpenAI. K3.1 Pro is a third cheaper than K3, but on price per token it matches Claude Sonnet 5.5 and GPT-6.1 Sol rather than undercutting them.

To use it in Kimi Code without paying per token, you need a Kimi membership. Kimi’s pricing page lists Kimi Code on Plus, Pro, Max and Ultra, at $180, $372, $948 and $1,908 a year, billed annually. Pro and above add Highspeed and the one-million-token context. For new members, Kimi’s membership guide says “the weekly quota limit is removed, and only the rolling 5-hour rate window remains,” though a monthly total still applies, and when it runs out “Kimi Code quota is frozen until the monthly quota resets or you upgrade.” It doesn’t publish how much either one allows.

OpenCode Go, OpenCode’s subscription to open models, includes K3.1 Pro too. On the $10-a-month Go plan, the model has a monthly limit of $15 of use at API prices, which OpenCode estimates at 147 requests every five hours, 368 a week or 735 a month. Go Plus, at $40, raises the limit to $60.

Five ways to use it in a coding agent

1. Kimi Code. Moonshot’s own agent comes as a CLI, a desktop app and a VS Code extension. Its docs install the CLI with curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash; run kimi, sign in with /login, and choose K3.1 Pro with /model. It needs the Plus membership or above.
2. Claude Code. Kimi’s guide points Claude Code at Kimi with a Kimi Code membership key: set ANTHROPIC_BASE_URL to https://api.kimi.ai/coding/, ANTHROPIC_API_KEY to the key, and ANTHROPIC_MODEL and Claude Code’s default model variables to k3.1-pro, or k3.1-pro[1m] for the one-million-token context. /status then shows Kimi’s address. Every request goes to Kimi instead of Anthropic, and thinking has to stay on: K3.1 Pro rejects requests without it.
3. Codex. Kimi’s Codex guide adds a provider to ~/.codex/config.toml with base_url = "https://api.moonshot.ai/v1" and wire_api = "responses", and reads a Kimi API key from KIMI_API_KEY. You pay per token.
4. OpenCode. Run /connect, choose Moonshot AI, enter a key from Kimi’s API platform, and pick the model with /models, as OpenCode’s provider guide describes. Models.dev, the catalog OpenCode takes its model lists from, lists Kimi K3.1 Pro under Moonshot AI and OpenCode Go.
5. Your own server. Once vLLM or SGLang is serving the weights, OpenCode can use them as a custom provider, with @ai-sdk/openai-compatible and your server’s address as options.baseURL.

Enjoy runs OpenCode beside Claude Code, Codex and the other labs’ agents. Add Moonshot AI to OpenCode once, with opencode auth login in a terminal, and Kimi K3.1 Pro appears in Enjoy’s model picker with the rest of OpenCode’s models; a server of your own goes in OpenCode’s config. Hand it a conversation when a Claude or ChatGPT plan runs low, or give it the same task to compare. Enjoy is a monthly subscription, with the first seven days free (pricing); Kimi bills you for the tokens, and Enjoy doesn’t add Kimi to a Claude or ChatGPT account.

The report that it learned from GPT-6 Astra

On October 28, 2026, The Information reported that Moonshot and DeepSeek “relied heavily” on OpenAI’s GPT-6 Astra to train their latest models, citing people with direct knowledge of the matter. The models named are Kimi K3.1 Pro and the October 17 update of DeepSeek’s V4.1 Flash. The report is paywalled; the details here are from Reuters’ and The Next Web’s accounts of it, and The Next Web says it “has not independently verified the report.”

Training one model on another’s answers is called distillation. It’s common practice, and the NSA, CISA and FBI advisory of September 8 calls it “a legitimate and useful technique in AI research.” What’s disputed is collecting a rival’s outputs at scale against its terms. That advisory accused six China-based companies, Moonshot AI and DeepSeek among them, of “aggressive, malicious, and targeted distillation activities at an industrial scale” against American models, Claude, GPT, Gemini and Grok included. China’s foreign ministry rejected it as unfounded.

It isn’t the first time Moonshot has been named. In a post on September 30, OpenAI said it had stopped a campaign, begun on July 1, to extract its models’ hidden reasoning, peaking on July 24 and 25 at 16,000 requests from over 4,000 users, and that “we attribute a core cluster of the activity to individuals associated with Moonshot AI, the developer of Kimi.” It added that “it is unclear whether all operators we observed during the relevant time period originated from a single actor,” and said it had shut the campaign down by July 28.

What the new report doesn’t establish:

• Moonshot’s side. Moonshot, DeepSeek and OpenAI didn’t respond to Reuters, and by November 4 we found no public comment from any of them. Moonshot’s announcement credits K3.1 Pro’s gains to “substantially larger post-training compute” and doesn’t mention distillation.
• How much, and how. The public accounts don’t say how many Astra outputs were used, how they were obtained, or how much of K3.1 Pro’s training they made up.
• Any change for users. We found no announced change to Kimi’s apps, Kimi Code or the Kimi API, and the weights came out on the promised day.

For most people, the practical question is the one it already was: whether K3.1 Pro does your work well enough at its price. If where a model’s training data came from matters to your company, ask the vendor before you build on it.

The creator brief nobody has authenticated

On launch day, an X creator, Elshayib, posted screenshots of what they called a Kimi campaign brief, Implicator reports. It offered the line “China just made Opus 5.5 free to use for everyone” and said “Please DO NOT put on the paid partnership label.” Implicator and RuntimeWire both say the screenshots haven’t been authenticated, and that it isn’t established who wrote the brief or whether Moonshot approved it. Moonshot hadn’t commented when Implicator reported on October 25, and we found no comment by November 4.

Either way, the line is wrong on its face. On the API, K3.1 Pro costs half what Opus 5.5 does, and free weights still need GPUs that rent for about $1,700 a day. Weigh glowing posts about K3.1 Pro accordingly.

What Moonshot hasn’t said

• What it takes to run, in memory or GPUs. The figures above come from vLLM, NVIDIA and Runpod.
• Highspeed’s weights, or whether Highspeed is a separate model at all.
• Who its certified inference partners are, at least on the model card.
• China’s draft export rules. K3.1 Pro is over the draft’s 2-trillion-parameter line, and we found nothing from Moonshot on whether it registered this release.
• How much each Kimi plan includes in Kimi Code. The membership guide describes the windows, not their size.
• Whether it trained on GPT-6 Astra. It hasn’t answered The Information’s report.
• Anything about the creator brief.

We’ll update this memo when Moonshot or OpenAI answers the report, or when a smaller build changes what it takes to run.