Kimi K3.1 Pro: Benchmarks, Pricing, and Open Weights

Kimi K3.1 Pro is free to download. Moonshot put the full weights on Hugging Face on November 4, and no open-weight model scores higher on the independent leaderboards right now.

It holds its own on coding, too. On Terminal-Bench 4.0, it sits less than a point behind Claude Opus 5.5. And Moonshot's API charges $2 per million input tokens, a third less than Kimi K3.

One thing to know up front: you won't run this on your laptop. It needs data center hardware, so most people will use the Kimi app or the API.

I read the model card, the license, and Moonshot's price pages to pull out what matters, including the license catch for bigger companies and the benchmarks where it still trails.

TL;DR:

• What it is: Moonshot AI's newest model, announced October 23, 2026, and a post-trained upgrade of Kimi K3 aimed at long coding jobs, 3D and CAD work, and financial research.
• Size: 2.8 trillion total parameters, 104 billion active per token, and a 1M-token context window.
• Open weights: released on Hugging Face on November 4 under an MIT-style license, with extra conditions only for model-hosting businesses over $20 million in revenue and very large consumer products.
• Price: $2 per million input tokens and $10 per million output tokens on the Kimi API, down from K3's $3 and $15.
• Highspeed version: 2 to 3 times faster by Moonshot's count, at $4 and $20, and available on the Kimi app's Pro plan and above.
• Where to use it: kimi.ai, the Kimi mobile apps, Kimi Work, Kimi Code, the Kimi API, Amazon Bedrock, and more than a dozen third-party hosts.
• How good it is: 65.7% on Terminal-Bench 4.0 by Moonshot's count, just behind Claude Opus 5.5 (66.4%), and 46 on Artificial Analysis's Intelligence Index, the top open-weight score but behind the closed leaders.

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is Moonshot AI's newest open-weight model, announced on October 23, 2026, as the first model in its K3.1 series. Moonshot's launch post calls it "our most capable 3T-class open-weight model to date."

It's a post-trained version of Kimi K3, the model it replaces.

It keeps the same 2.8 trillion parameters, and Moonshot spent "substantially larger post-training compute" on reinforcement learning across hundreds of task types, from coding to finance.

The specs, from Moonshot's model card and API docs:

K3.1 Pro swaps the full-attention layers K3 used for a sparse attention design borrowed from DeepSeek, which Moonshot says handles long inputs more efficiently.

Kimi K3.1 Pro open weights: what shipped

Moonshot released Kimi K3.1 Pro's full weights on November 4, 2026, hitting the "by Nov 4" date it set at launch. Its announcement on X linked straight to the Hugging Face repo, which is public and doesn't require an access request.

Artificial Analysis has since moved it into its open-weights category, with a "commercial use restricted" label because of the license conditions below.

The repo holds the full checkpoint as 96 safetensors files, about 1.56 TB in total.

The model's expert layers ship in MXFP4, a 4-bit format, and everything else in BF16, a 16-bit format. That squeezes 2.8 trillion parameters into far less space than a standard 16-bit release would need (about 5.6 TB). It's still far more memory than any desktop computer has.

What the license allows

Kimi K3.1 Pro uses its own "Kimi K3.1 Pro License," and the license file reads like the MIT license with two extra conditions bolted on for big model-hosting businesses and very large products.

Here's what it says:

• Allowed for everyone: using, copying, modifying, fine-tuning, distributing, and selling the model or products built on it, free of charge.
• Condition 1, model-as-a-service: this covers businesses selling model access, such as an API where customers control the inputs or training data. If your company runs one and makes more than $20 million in revenue over any 12 months, you need a separate agreement with Moonshot before using K3.1 Pro commercially.
• Condition 2, big consumer products: if a commercial product built on it has more than 100 million monthly active users or more than $20 million in monthly revenue, it has to display "Kimi K3.1-Pro" prominently in the interface.
• Exempt: internal use that doesn't expose the model, its outputs, or its capabilities to outside parties.

For almost every team, that means you can run it, fine-tune it, and ship products on it without asking anyone. The conditions only bite for large inference providers and products at a scale most companies never reach.

It's a similar shape to the license Mistral Large 4 shipped under in October, which also singles out model-as-a-service businesses, though Moonshot adds a revenue threshold before the condition applies.

What hardware you'll need

Kimi K3.1 Pro needs a full data center node to run at full quality. The vLLM recipe lists tested setups for the official MXFP4 checkpoint, including these:

Every option is data center hardware with at least eight accelerators. On one B300 node, vLLM's fastest setup read in a 340,000-token prompt in about 10 seconds before it started answering.

So, can you run it locally? Not on a laptop, a gaming PC, or a high-end Mac (at least not without squeezing it into a much smaller quantized file, which costs some quality).

For most people, "open weights" here means your company can run it on its own servers or pick the cloud host it trusts.

Which tools can run it

K3.1 Pro had day-one support in the main open-source serving engines. The model card recommends:

• vLLM: recipes for NVIDIA, AMD, Zhenwu PPUs, and Google TPUs, starting at vLLM 0.34.0
• SGLang: a cookbook entry with its own launch settings
• TokenSpeed: a newer inference engine with its own K3.1 Pro recipe

Two quirks matter if you build on it. The model always thinks (you can set reasoning effort to low, high, or max, with max as the default), and Moonshot's docs say multi-turn chats must send back its full previous reply, including the reasoning and the tool calls.

Kimi K3.1 Pro pricing

Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens on the Kimi API, with cached input at $0.20. That's a third cheaper than Kimi K3, which still lists at $3 and $15.

Here's the full rate card from Moonshot's pricing page, checked on November 9:

A few details that change the math:

• Caching is automatic: Repeated prompt prefixes get cached for five minutes by default, and you can pay double the write price for a one-hour cache if your requests come further apart.
• Batch is 40% off: Jobs that don't need an instant answer run through the Batch API at 60% of the standard price.
• No long-context surcharge: The price is flat across the full 1M-token window.
• You need a $1 top-up first: Moonshot unlocks K3.1 Pro after your first successful payment, and your total top-ups set your rate limits.

Our Kimi K3 pricing guide covers how the cache-write charges and app plans work in more depth. The same mechanics apply to K3.1 Pro, just at the lower rates above.

What is Kimi K3.1 Pro Highspeed?

Kimi K3.1 Pro Highspeed is a faster version of the same model that Moonshot says runs at "2-3× the speed of K3.1 Pro." It costs exactly double on the API ($4 input, $20 output), though cached input stays at $0.20.

It's the one to pick for chat-style back-and-forth or coding sessions where you're watching the cursor. For overnight batch jobs, the standard model at half the price makes more sense.

In the Kimi app, Highspeed is only on the Pro plan and above.

How to get Kimi K3.1 Pro

Kimi K3.1 Pro is live everywhere Moonshot sells Kimi: the web app at kimi.ai, the Kimi mobile apps (iOS, Android, and HarmonyOS), the Kimi Work desktop app (version 3.1.0 or later), Kimi Code, and the API.

Here's what the app plans cost, from Moonshot's membership page:

Max and Ultra also unlock the full 1M-token context in chat. And a heads-up from the same page: Moonshot warns that "K3.1 Pro requests may queue or fail at launch," so expect some waiting in the first weeks.

The other access points:

• Kimi Code: run it and switch models with the /model command.
• Kimi API: pick kimi-k3.1-pro or kimi-k3.1-pro-highspeed. The API speaks both the OpenAI and Anthropic formats, and Zero Data Retention is available.
• Kimi Enterprise: admins have to switch K3.1 Pro on, since access was off by default at launch.
• Cloud platforms: Amazon Bedrock added K3.1 Pro (Highspeed) on November 6, in every region Bedrock covers through cross-Region inference, and NVIDIA lists it in its API catalog under trial terms.

Third-party hosts

Seventeen providers, Moonshot included, serve K3.1 Pro through OpenRouter, across 24 endpoints. Eight of those endpoints match Moonshot's $2 and $10, several list less (PrimaLabs and InferenceNet charge $1.80 and $9), and a few charge more, including Alibaba at $2.30 and $11.50.

Third-party hosts are also where the speed is. Artificial Analysis clocks Moonshot's own API at about 32 tokens per second and Highspeed at about 61, while the fastest third-party endpoint it tracks (LithosAI) runs at about 249 tokens per second for $6 and $30 per million tokens.

Kimi K3.1 Pro vs. K3: what changed

K3.1 Pro keeps K3's 2.8 trillion parameters, adds new attention layers, and gets much more training on real tasks, and it costs less. Moonshot frames it as a post-training upgrade, so the gains show up mostly in coding, finance, and factual accuracy.

Here's what changed:

• More reinforcement learning: Moonshot optimized every layer of its training stack so it could run reinforcement learning across hundreds of task domains at once.
• Far fewer made-up facts: on Moonshot's internal Hallu Bench Hard, the share of false claims fell from 23.4% for K3 to 8.3%, and answers with at least one factual error became less than half as common (42.4% of K3's rate).
• Better finance work: +7.2 points on Moonshot's internal financial modeling test, +11.6 points on financial research, and about 4.2 times K3's score on a simulated business-management test.
• Better long coding jobs: 61.2% on Kimi Code Bench 3.0 Long-horizon, up from K3's 55.95%.
• New attention layers: sparse DeepSeek-style attention replaces K3's full-attention layers for long inputs.
• Lower price: $2 and $10 per million tokens, down from $3 and $15.

Moonshot is careful to say the hallucination test is a stress test "deliberately crafted to probe situations where models are prone to errors," so those rates aren't what you'd see in everyday use.

On the benchmarks that appear on both model cards, here's how the two compare:

All of these are Moonshot's own numbers. Treat the DeepSWE jump with some care: Moonshot ran K3 inside its own Kimi Code tool and K3.1 Pro inside Claude Code, so part of the gap could come from the tool each model ran in.

Independent tests show a bigger jump than Moonshot's own table.

Vals AI puts K3.1 Pro at 62.52% on its Vals Index, up from 52.95% for K3, and on Terminal-Bench 4.0 it went from 17.17% to 63.64%. Artificial Analysis doesn't currently show an Intelligence Index score for K3 at max effort, so there's no clean before-and-after there.

If you already run K3, there isn't much reason to stay. K3.1 Pro uses the same 1M context and the same API format, and it's cheaper per token.

Kimi K3.1 Pro benchmarks: Moonshot's numbers vs. independent tests

Kimi K3.1 Pro is the strongest open-weight model on the independent leaderboards, and it trails the top closed models overall even though Moonshot's own chart puts it within a point of Claude Opus 5.5 on Terminal-Bench 4.0.

Moonshot's numbers

Moonshot's model card compares K3.1 Pro with Claude Opus 5.5, GPT-6 Astra, Claude Fable 5.1, DeepSeek V4.1 Flash, and GLM-5.3 on nearly 50 tests. Here's a selection:

All scores are self-reported at max reasoning effort. Moonshot ran K3.1 Pro inside Anthropic's Claude Code harness for most coding tests and took most rival scores from their makers' reports and official leaderboards, though it ran some comparisons itself, including OfficeQA Pro v2.

What independent testing shows

On Artificial Analysis, K3.1 Pro scores 46 on the Intelligence Index, the top score among open-weight models as of November 9.

That's well ahead of the next open-weight models, DeepSeek V4.1 Flash (33) and GLM-5.3 (32), but 11 to 14 points behind the closed leaders: Gemini 4 Carbon (60), Claude Opus 5.5 (59), and GPT-6 Astra (57).

Artificial Analysis puts its cost at $3.68 per Intelligence Index task, helped by the $2 and $10 list price.

Vals AI tells a similar story. K3.1 Pro ranks #9 of 34 on the Vals Index at 62.52%, the best open-weight result, with Opus 5.5 at 70.18%.

On Terminal-Bench 4.0, the outside numbers bracket Moonshot's:

• Artificial Analysis: 67.2%, the top score on that leaderboard, ahead of Gemini 4 Carbon (65.2%) and Claude Opus 5.5 (59.6%)
• Moonshot: 65.7%
• Vals AI: 63.64%, #3 of 61 models and about 1.5 points behind Opus 5.5

So Moonshot's Terminal-Bench number holds up, while the broad indexes still put K3.1 Pro a tier below Gemini 4 Carbon and Opus 5.5.

Where it falls behind

Moonshot's own table and Artificial Analysis's tests both show weak spots, and they're worth knowing before you switch:

• Hard reasoning without tools: 39.4% on HLE-Diamond against 60.9% for Opus 5.5 and 65.8% for GPT-6 Astra.
• Science coding: 38.1% on Terminal-Bench-Science, 30 points behind GPT-6 Astra's 68.1%.
• Building full programs: 31.5% on ProgramBench, about half of Opus 5.5's 65.0%.
• Office work: an AA-Briefcase Elo of 1,678, behind Opus 5.5's 1,803.
• Speed: Artificial Analysis measured about 32 tokens per second on Moonshot's API and over three minutes before the first answer token, with reasoning set to max by default.
• Hallucinations on outside tests: Artificial Analysis's AA-Omniscience test, which punishes confident wrong answers, puts its hallucination rate at 37.3%, far above the 8.3% from Moonshot's in-house test. K3 scored 76.6% here, but only at low effort, so treat that comparison loosely.

‍

{{templates}}

‍

What Kimi K3.1 Pro is best at

Moonshot leads with three jobs: long-running coding, 3D and CAD work, and financial research. It also made an unusual call on cybersecurity for the open release.

Long-horizon coding

K3.1 Pro is built for coding jobs that run for hours, and it leads Moonshot's table on SWE Marathon v1.1 (51.7%), a test of very long software tasks, ahead of Claude Opus 5.5 (46.3%) and GPT-6 Astra (42.5%).

In Moonshot's showcase example, working on top of an open-source model of the human body (296 skeletal parts and 402 muscles), K3.1 Pro built a walking simulator across 167 missions and 420 separate reviews.

When tweaking the reward stopped helping the arms swing properly, it went digging in the simulator itself, found and fixed 36 errors in the shoulder-muscle math, then retrained.

The simulated human went from collapsing within seconds to walking for 11.2 seconds straight.

(Anyone who's watched an AI agent "fix" a failing test by deleting it will appreciate that it went and fixed the physics.)

Moonshot says it also works across game engines like Unity, Unreal Engine, and Godot, and that its front-end pages and Three.js scenes show stronger composition, typography, animation, and polish.

3D and CAD work

K3.1 Pro scores 92.7% on BenchCAD with tools, a test that has models rebuild 3D objects from multi-view images by writing CAD code. That's close to Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%), and well ahead of other open models like DeepSeek V4.1 Flash (81.6%).

Moonshot says it's strongest at parametric modeling: mechanical parts, architectural models, configurable products, and custom furniture. Without MCP connectors, it can write and run scripts for Blender and CAD software to build and edit models.

Moonshot shipped a consumer feature alongside it, too. 3D Make in the Kimi app lets you preview and edit 3D models, then download them ready for 3D printing.

Financial research

K3.1 Pro posts the top score in Moonshot's table on OfficeQA Pro v2 (64.4%), which tests analysis across real financial documents, ahead of Opus 5.5 (62.2%) and GPT-6 Astra (54.4%).

Moonshot targeted factual accuracy because finance work punishes made-up numbers, and on its own stress test, K3.1 Pro's 8.3% claim-level error rate beats Opus 5.5 (15.0%), though it trails GPT-6 Astra (5.5%).

Claude Fable 5.1 edges it on Frontier Finance, which covers full investment workflows (60.8% to 58.6%).

The cybersecurity choices behind the open release

Moonshot deliberately made K3.1 Pro worse at hacking before releasing the weights.

While training, Moonshot noticed the model's vulnerability research skills kept climbing. It traced that to cyber-related content mixed into its general training data and its reinforcement-learning tasks.

So for K3.1 Pro, which it planned to release as open weights, it removed that data from both stages while keeping the general coding tasks.

The effect is measurable on Moonshot's internal exploit test, xdev-bench-hard:

• Without the data removal: 90.5% of tasks solved
• The released K3.1 Pro: 81.0%

Moonshot says the released model's cyber skills still trail frontier proprietary models that have been public for at least three months.

Its stated rule is to avoid releasing open weights that push the cyber frontier, then raise the ceiling gradually so defenders can audit their own code, including old systems that rarely get a security review.

The public cyber benchmarks show the same gap. K3.1 Pro scores 20.7% on ExploitGym against 34.5% for Opus 5.5 and 42.4% for GPT-6 Astra. If security research is your main use case, the closed models are still ahead.

Should you use Kimi K3.1 Pro?

Use Kimi K3.1 Pro if you want the best open-weight model available, or near-frontier coding at $2 and $10 per million tokens. Pass on it if you need the absolute top score, fast replies, or the strongest security research.

Choose Kimi K3.1 Pro if you:

• Want to self-host a top model on your own hardware, for privacy, compliance, or cost control
• Run long coding agents and care about cost per task more than speed
• Do 3D, CAD, or parametric design work
• Work with financial documents and want a model Moonshot tuned for factual accuracy
• Already use Kimi K3 (it's cheaper and better)

Look elsewhere if you:

• Need the top overall model today (Gemini 4 Carbon and Claude Opus 5.5 sit near the top of both independent indexes)
• Want quick answers in chat, since it's slow at max effort unless you pay for Highspeed or a fast third-party host
• Do vulnerability research, where Moonshot deliberately held its cyber skills back
• Want something you can run on a laptop (smaller open models are the realistic pick there)

For a closer look at two of the closed models it's chasing, see our Opus 5.5 vs. GPT-6 Astra comparison.

And if you're tired of changing tools every time a new model tops a chart, Lindy, an AI teammate that works in Slack, lets you pick your model for each task from most major models and move to a new one when it pulls ahead.

‍

{{cta}}

‍

FAQ

Is Kimi K3.1 Pro open source?

Kimi K3.1 Pro is open-weight: the full weights have been free to download from Hugging Face since November 4, 2026. It uses Moonshot's Kimi K3.1 Pro License, which works like MIT for most users but adds conditions for model-hosting businesses whose revenue tops $20 million over any 12 months, and for products with over 100 million monthly users or over $20 million in monthly revenue.

How much does Kimi K3.1 Pro cost?

Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens on the Kimi API, with cached input at $0.20 and batch jobs at 60% of the standard price. Kimi K3.1 Pro Highspeed costs $4 and $20. In the Kimi app, paid plans run from $19 to $199 a month, and the weights are free if you host the model yourself.

Is Kimi K3.1 Pro better than Claude Opus 5.5?

No, Kimi K3.1 Pro trails Claude Opus 5.5 overall, though it matches or beats it on some coding tests. Moonshot reports 65.7% on Terminal-Bench 4.0 against Opus 5.5's 66.4%, and K3.1 Pro leads on SWE Marathon and OfficeQA Pro v2. On Artificial Analysis's Intelligence Index, Opus 5.5 scores 59 to K3.1 Pro's 46, and Vals AI shows a similar gap, though K3.1 Pro is open-weight, so you can host it yourself.

What is Kimi K3.1 Pro Highspeed?

Kimi K3.1 Pro Highspeed is a faster version of Kimi K3.1 Pro that Moonshot says runs at 2 to 3 times the speed of the standard model. On the API it's called kimi-k3.1-pro-highspeed and costs $4 per million input tokens and $20 per million output tokens. In the Kimi app, it's included on the Pro, Max, and Ultra plans.

Can I run Kimi K3.1 Pro locally?

Only on data center hardware, such as your company's own GPU servers. The official weights take about 1.56 TB, and the tested setups in vLLM's recipe include a full eight-GPU NVIDIA B300 node, two GB300 trays, an AMD MI355X node, Zhenwu PPU nodes, or a 16-chip Google TPU slice. vLLM, SGLang, and TokenSpeed all support it.

How is Kimi K3.1 Pro different from Kimi K3?

Kimi K3.1 Pro is Kimi K3 with much more reinforcement learning, new attention layers, and a lower price. Moonshot reports its hallucination rate on a hard internal test fell from 23.4% to 8.3%, its finance and long-horizon coding scores rose, and its API price dropped from $3 and $15 to $2 and $10 per million tokens. Both have 2.8 trillion parameters and a 1M-token context window.