Try Kimi at kimi.com →

Kimi K3.1 Pro
Features, Access & Use Cases

Kimi K3.1 Pro is Moonshot AI's newest and most capable model: an open-weight, 2.8-trillion-parameter model that reads up to a million tokens at once and understands text, images and video. This guide explains what it can do, how it compares with Kimi K3, every way to use it, what it costs and where it shines.

Open weights 1M-token context API $2 / $10 per 1M
K3.1PRO PARAMETERS 0.7T 1.4T 2.1T 2.8T CONTEXT 256K 512K 768K 1M tokens ACTIVE PER TOKEN 26B 52B 78B 104B Long context Long context Vision Vision Agents Agents Coding Coding Finance Finance 3D & CAD 3D & CAD Open weights · $2 / $10 per 1M
AT A GLANCE

Kimi K3.1 Pro at a glance

2.8Tparameters

104B active per token across 896 experts.

1Mtoken context

1,048,576 tokens with efficient sparse attention.

$2 / $10per 1M tokens

API input and output. Cached input $0.20.

Openweights

Free to download from Hugging Face.

Full specifications

  • DeveloperMoonshot AI
  • ReleasedOctober 2026 (open weights 4 November 2026)
  • ArchitectureMixture of experts: 896 experts, 16 active per token, 93 layers
  • Parameters2.8 trillion total, 104 billion active per token
  • Context window1,048,576 tokens
  • AttentionSparse gated attention in 24 of 93 layers for long inputs
  • InputsText, code, images and video (MoonViT-V2 vision encoder)
  • Thinking effortLow, High or Max
  • API model IDkimi-k3.1-pro
  • API price$2 input · $10 output · $0.20 cached, per 1M tokens
  • WeightsAbout 1.56 TB on Hugging Face (MXFP4 weights, MXFP8 activations)
  • LicenceKimi K3.1 Pro License (permissive, with terms for very large companies)

Specifications and prices from Moonshot AI's launch materials and API documentation.

UPGRADE

What's new compared with Kimi K3

K3.1 Pro builds on Kimi K3, released in July 2026, with the same size and context but better efficiency, accuracy and new 3D skills, at a lower price.

Kimi K3Kimi K3.1 Pro
API input / output$3 / $15$2 / $10
Cached input$0.30$0.20
Long-context attentionFull attentionSparse attention in 24 layers
False statements (finance tests)23.4%8.3%
3D, CAD and game enginesBasicBlender, CAD, Unity, Unreal, Godot, 3D Make
Open weightsYes (July 2026)Yes (4 November 2026)
Highspeed version—$4 / $20, 2–3× faster
KIMI K3.1 PRO · FEATURES Context window 250K 500K 750K PDF XLS PNG DOC MP4 CODE 1M tokens reads it all at once Thinking effort Low Low High High Max Max Reasoning depth Understands Text Images Video Code False statements (finance) Kimi K3 K3.1 Pro 23.4% 8.3% −65% errors Moonshot's internal finance tests

The 1M-token context, thinking effort, multimodal input and the drop in false statements.

FEATURES

Key features

01

1M-token context

Holds a whole codebase or stack of long reports at once. Sparse attention keeps long inputs efficient.

02

Native vision

Reads screenshots, charts, scanned pages, photos and video alongside text.

03

Adjustable thinking

Low for quick answers, High for everyday work, Max for the hardest problems.

04

Agents and tool use

Plans multi-step tasks, browses, uses a computer and calls tools, scoring 79.6 on OSWorld 2.1.

05

Strong coding

65.7 on Terminal-Bench 4.0, close to the best closed models, and the engine behind Kimi Code's hardest jobs.

06

3D, CAD and games

Writes Blender and CAD scripts, builds games in Unity, Unreal and Godot, and powers Kimi 3D Make.

07

More accurate finance

Moonshot reports 14–20% gains on finance tests and far fewer false statements than K3.

08

Open weights

Download, fine-tune and run it yourself under a permissive licence.

09

Lower price

A third cheaper than K3 on the API, with cached input at $0.20 per million tokens.

UNDER THE HOOD

How K3.1 Pro works

INSIDE KIMI K3.1 PRO Mixture of experts: only the experts each word needs INPUT ROUTER Kimi "Kimi" → 6 experts shown reads "reads" → 6 experts shown long "long" → 6 experts shown files "files" → 6 experts shown 56 of 896 experts shown 896experts 16active per word 104Bof 2.8T params used

Each word goes to a few of 896 experts, so only 104B of 2.8T parameters work at a time.

  • Mixture of experts. Instead of using all 2.8 trillion parameters for every word, a router picks 16 of 896 specialist "experts". That keeps a huge model fast and affordable.
  • Sparse attention for long inputs. 24 of its 93 layers use a gated sparse attention design, so it can work across a million tokens without slowing to a crawl.
  • Built-in vision. A 401-million-parameter vision encoder, MoonViT-V2, lets it read images and video natively.
  • Compact weights. MXFP4 weights and MXFP8 activations shrink the download to about 1.56 TB.
  • Safer by design. Moonshot removed cybersecurity exploit data from training, so it's deliberately weaker at writing exploits.
BENCHMARKS

How K3.1 Pro compares

Coding in a terminal

Terminal-Bench 4.0 score out of 100. Higher is better.

Claude Opus 5.566.4
Kimi K3.1 Pro65.7
GPT-6 Astra58.2

More K3.1 Pro scores

  • GPQA-Diamond (graduate-level science)94.2
  • BenchCAD (3D from CAD code, with tools)92.7
  • BrowseComp (web research)92.1
  • OSWorld 2.1 (using a computer)79.6

All scores are Moonshot's own, from the K3.1 Pro launch materials. Independent results may differ; treat small gaps as a tie.

ACCESS

How to access Kimi K3.1 Pro

WhereHowCost
Kimi app & kimi.comChoose the newest model in chatUses monthly credits
Kimi WorkDesktop agent for Windows and MacUses monthly credits
Kimi CodeSwitch model for harder coding jobsKimi Code allowance
Kimi APIModel ID kimi-k3.1-pro$2 / $10 per 1M
Hugging FaceDownload open weights, self-hostFree (your hardware)
Cloud partnersFor example DatabricksPartner pricing

Chat with the newest model draws on credits; the free plan has a small allowance. Plans and credits

WHERE TO USE K3.1 PRO Kimi app & webChat and tools Kimi WorkDesktop agent Kimi CodeTerminal & IDE Kimi APIkimi-k3.1-pro Hugging FaceOpen weights DatabricksCloud partner K3.1PRO

Six ways to use K3.1 Pro, from the Kimi app to your own servers.

SPEED

Standard vs Highspeed

K3.1 ProK3.1 Pro Highspeed
API input / output$2 / $10 per 1M$4 / $20 per 1M
SpeedStandard2–3× faster replies (Moonshot's figure)
What gets faster—Model output only, not tool calls or scripts
In Kimi CodeStandard quotaAbout 2× quota, from Allegretto
Best forMost work, batch jobs, cost controlInteractive coding where you wait on every reply
USE CASES

What Kimi K3.1 Pro is best at

  • Research across many documents. The 1M context means whole reports, papers or filings can be compared in one go, with citations.
  • Large codebase work. Read an entire repo, refactor across files and fix failing tests in Kimi Code.
  • Financial analysis. Build models from filings with far fewer false statements than K3.
  • 3D design and CAD. Generate printable models, Blender and CAD scripts, or game prototypes.
  • Long agent tasks. Plan and carry out multi-step jobs with browsing and tools.
Research Research Coding Coding Finance Finance 3D & CAD 3D & CAD “Compare these 6 reports and cite every claim”K123456 reports · 41 citations “Refactor the auth module and fix failing tests”K- session = getSession(req)+ session = await getSession(req)- if (!user) return+ if (!user) throw new AuthError()✓ 42 tests passed “Build a 3-year model from these filings”KFalse statements ↓ 23.4% → 8.3% “Design a printable phone stand, 10 cm tall”K Export for 3D print POWERED BYKimi K3.1 Pro · 1M context · thinking: Max

Research, coding, finance and 3D: four jobs where K3.1 Pro stands out.

Prompts to try with K3.1 Pro

Research

Literature and market reviews

Read these [number] reports and compare their main findings in a table. Cite the page for each point and list where they disagree.

Coding

Large-codebase changes

Read this repository, then refactor [module] for readability without changing behaviour. Run the tests and explain each change.

Finance

Models from filings

Using the attached annual reports, build a three-year revenue and margin model in a table. Flag every assumption and cite its source.

3D & CAD

Printable designs

Design a 3D [object] about [size] cm tall for 3D printing, with walls at least 2 mm thick. Explain how to export it.

Vision

Screens to code

Turn the attached screenshot into a responsive HTML and CSS page. Match the layout and keep everything in one file.

Agents

Multi-step jobs

Plan and carry out this task step by step: [task]. Show the plan first and wait for my OK before making changes.

Set thinking to Max for the hardest of these. More Kimi prompts

DEVELOPERS

API quick start

Python · OpenAI SDK
# pip install --upgrade openai from openai import OpenAI client = OpenAI( api_key="YOUR_KIMI_API_KEY", base_url="https://api.moonshot.ai/v1", ) r = client.chat.completions.create( model="kimi-k3.1-pro", reasoning_effort="max", # low | high | max messages=[{"role": "user", "content": "Review this contract and list the risks."}], ) print(r.choices[0].message.content)

Get started

  1. Create an account at platform.kimi.ai and top up from $1.
  2. Create an API key and keep it secret.
  3. Point any OpenAI-compatible SDK at https://api.moonshot.ai/v1.
  4. Set the model to kimi-k3.1-pro and choose reasoning_effort.

Thinking counts as output tokens. Cached input costs $0.20 per million. Kimi API guide and cost calculator

OPEN WEIGHTS

Self-hosting and licence

Running it yourself

  • Download: public, ungated weights on Hugging Face (about 1.56 TB).
  • Serve: Moonshot provides recipes for vLLM, SGLang and TokenSpeed.
  • Hardware: roughly eight or more B300-class GPUs for the full model.
  • Why: keep data in-house, fine-tune, or avoid per-token costs at very high volume.

Kimi K3.1 Pro on Hugging Face

The Kimi K3.1 Pro License

  • Permissive: use, change, fine-tune and sell products built on it.
  • Large resellers: services earning over $20M a year from offering the model need a separate agreement.
  • Very large products: over 100M monthly users or $20M monthly revenue must display the "Kimi K3.1-Pro" name.

This is a summary, not legal advice. Read the full licence before commercial use.

GOOD TO KNOW

Limitations

  • Benchmarks are Moonshot's own. Independent tests may rank it differently; try it on your real tasks.
  • Not the fastest by default. Max thinking on long inputs takes time. Highspeed helps, at double the price.
  • Data location. Moonshot's own service runs through China. Use self-hosting or a cloud partner if that matters.
  • Self-hosting is expensive. The weights are free, but the hardware is not.
  • It can still be wrong. Fewer false statements is not zero. Check important facts and numbers.
FAQ

Kimi K3.1 Pro FAQ

Kimi K3.1 Pro is Moonshot AI's newest flagship model, launched in October 2026 with open weights released on Hugging Face on 4 November 2026. It is a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters per token and a 1,048,576-token context window.

You can try it in the Kimi app, where chat with the newest model draws on your monthly credits, including a small allowance on the free plan. The open weights are free to download, and the API costs $2 per million input tokens and $10 per million output tokens.

K3.1 Pro is a third cheaper on the API, handles long inputs more efficiently with sparse attention, makes far fewer false statements on finance tasks (8.3% versus 23.4% in Moonshot's tests), and adds strong 3D, CAD and game-engine skills.

Use it in the Kimi app and website, Kimi Work, Kimi Code or the Kimi API with the model ID kimi-k3.1-pro. You can also download the open weights from Hugging Face or use cloud partners such as Databricks.

Highspeed serves the same model two to three times faster, according to Moonshot, at $4 per million input tokens and $20 per million output tokens. It speeds up the model's replies, not tool calls, and uses about twice the quota in Kimi Code.

Yes, but it needs data-centre hardware. The weights are about 1.56 TB, and running the full model takes roughly eight or more B300-class GPUs. Moonshot provides recipes for vLLM, SGLang and TokenSpeed.

Yes. The Kimi K3.1 Pro License allows commercial use. Services reselling the model with over $20 million in yearly revenue need a separate agreement, and products with over 100 million monthly users or $20 million monthly revenue must display the Kimi K3.1-Pro name.

Yes. It scores 65.7 on Terminal-Bench 4.0 in Moonshot's tests, close to Claude Opus 5.5 at 66.4 and ahead of GPT-6 Astra at 58.2, and it powers the hardest jobs in Kimi Code.

Try Kimi K3.1 Pro

Use it in the Kimi app, build with the API, or download the open weights.