What Kimi K2 does

Kimi K2 is an open-source artificial intelligence model portfolio developed by Moonshot AI, a Beijing-based AI startup founded in 2023. The portfolio centers on a proprietary mixture-of-experts (MoE) large language model with 1 trillion total parameters and 32 billion activated parameters, pre-trained on 15.5 trillion tokens using the company's novel MuonClip Optimizer, which is credited with achieving zero training instability. Active variants include Kimi K2 (original), Kimi K2 Thinking (reasoning-focused, released November 2025), Kimi K2.5 (multimodal vision language model, January 2026), and Kimi K2.6 (June 2026), with Kimi-K2-Base and Kimi-K2-Instruct checkpoints publicly available on HuggingFace.

The architecture is engineered for agentic intelligence — autonomous tool use, multi-step reasoning, and parallel orchestration of up to 100 sub-agents — and supports text, image, and video modalities. Benchmark results include 71.3% on SWE-Bench Verified, 44.9% on Humanity's Last Exam with tool assistance (surpassing GPT-5's 41.7%), and 83.1% on LiveCodeBench v6. Distribution spans self-deployment via open-source, the official Moonshot AI API (platform.moonshot.ai), NVIDIA GPU-accelerated endpoints on build.nvidia.com, AgentCard.org runtime infrastructure, and third-party platforms such as kimik2.com.

The revenue model combines usage-based API credit consumption, tiered credit package subscriptions (Starter $9.9 / Pro $49.9 / Ultimate $99.9 one-time), and enterprise licensing arrangements such as Cursor's Composer 2, which is fine-tuned on K2.5 and licensed through Fireworks AI. Customer segments span individual developers and indie hackers seeking cost-effective coding AI, AI startups integrating open-source models into products (Cursor, Perplexity), and large enterprise technology companies deploying K2.5 in internal programming tools (Alibaba, Tencent, ByteDance, Cloudflare). Following the January 2026 K2.5 release, Stripe-tracked individual subscription orders surged 8,280% month-over-month, and K2.5 became the most-called model on the OpenRouter routing platform.

Name

Kimi K2

Legal name

Moonshot AI

Website

https://kimik2.com

Company type

Private

Founded year

2023

Operating status

Operating

Short description

Kimi K2 is an open-source AI model portfolio developed by Moonshot AI featuring a 1-trillion-parameter mixture-of-experts architecture designed for agentic intelligence, coding, and multimodal tasks, serving developers, AI startups, and enterprise technology companies globally.

Ownership category

akta.pro rank

Product category

Large Language Models

NAICS

Software Publishers (5132), Software Publishers (513210)

SIC

Services-Computer Integrated Systems Design (7373)

akta.pro primary industry

End-to-End Enterprise AI Platforms (MLOps & Model Lifecycle Management) (HDAEANAA)

Keywords

• Open-source AI models
• Large language models
• Mixture-of-experts AI
• Agentic AI platform
• Multimodal foundation models

Headquarters

HQ city

Beijing

HQ country

China

HQ region

Asia

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Technology or R&D, Infrastructure, Personnel, Marketing or Sales, Operations

Revenue model

1. API Access Credits: Users purchase credits to access the Kimi K2 AI model through the platform. Credits are consumed per conversation or task, with different tier packages offering varying credit amounts and validity periods.
2. Subscription Credit Packages: One-time purchase credit packages with different tiers (Starter, Pro, Ultimate) offering 400-6000 credits with validity periods from 1 month to 1 year.
3. Enterprise Licensing: Enterprise licensing arrangements where companies like Cursor legally license the K2.5 model through intermediaries like Fireworks AI for integration into their products.

Pricing tiers

Distribution channels5 records

Marketing channels4 records

Core offering

Kimi K2, developed by Moonshot AI, is an open-source AI model portfolio built on a mixture-of-experts architecture with 1 trillion total parameters (32 billion activated). It provides language and multimodal AI capabilities—chat, reasoning, coding, mathematics, and agentic tool use—delivered via open-source checkpoints on GitHub and HuggingFace, an official API on platform.moonshot.ai, and NVIDIA GPU-accelerated endpoints.

Product overview

Kimi K2 is an open-source AI model portfolio developed by Moonshot AI featuring a mixture-of-experts architecture with 1 trillion total parameters. The core offerings include Kimi K2 (original model), Kimi K2.5 (multimodal vision language model), Kimi K2.6 (latest enhanced version), and Kimi K2 Thinking (reasoning-focused engine). Open-source variants include Kimi-K2-Base for fine-tuning and Kimi-K2-Instruct for general use, both available on HuggingFace. The platform (kimik2.com) provides web access to all models plus integrated access to DeepSeek, Claude, GPT, Gemini, Qwen, and Grok. Additional offerings include NanoBanana Pro for image generation and API access via platform.moonshot.ai. The products serve diverse use cases from coding and reasoning to multimodal agentic workflows.

Differentiator

Problem solved

Functional benefit

Brands

• Kimi-K2-Base: Open-source base model for fine-tuning, part of the Kimi K2 family with 1 trillion total parameters and 32 billion activated parameters.
• Kimi-K2-Instruct
• Kimi Claw

Products and services

• Kimi K2 State-of-the-art AI model leveraging mixture-of-experts architecture with 1 trillion total parameters and 32 billion activated parameters, designed for agentic intelligence with tool use, reasoning, and autonomous problem-solving capabilities.
• Kimi K2.5 Open multimodal vision language model featuring 1 trillion total parameters with 32.86 billion active parameters and 3.2% activation rate, designed for agentic AI workflows, chat, reasoning, coding, and mathematics, processing text, images, and video.
• Kimi K2.6 Latest version of Kimi K2 featuring enhanced agentic capabilities, available as a thinking model optimized for planning, exploring, and solving complex problems through deliberate reasoning steps.
• Kimi K2 Thinking Open-source intelligence engine built for real reasoning capable of planning, exploring, and solving through hundreds of deliberate steps, blending tool use with deep analytical thinking for state-of-the-art performance across search, coding, and complex problem-solving.
• Kimi-K2-Base Open-source base model version of Kimi K2 for fine-tuning, available on HuggingFace for organizations to customize for domain-specific applications.
• Kimi-K2-Instruct Open-source instruct model ready for general chat and agentic tasks, available on HuggingFace for community use and development.
• Kimi K2 API API access through platform.moonshot.ai enabling developers to integrate Kimi K2 capabilities into their applications, supporting coding, reasoning, and agentic tasks.
• NanoBanana Pro Image generation tool on the Kimi K2 platform supporting text-to-image and image-to-image generation, available as part of Pro subscription tier.

Quantifiable outcome

• 4.5x efficiency improvement with multi-agent orchestration (K2.5 managing 100 sub-agents)
• +4 more outcomes

Named customers7 records

Segments4 records

Ideal customer profiles4 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration12 records

AI capability11 records

Feature5 records

Partnerships

Six partnerships are on record, tiered minor, core and flagship.

• Huawei CloudminorStrategic or Co-development Partner · 8 June 2026Huawei Cloud partnered with Kimi through their AI Model Partner Program at INSPIRE 2026, alongside Zhipu AI, DeepSeek, MiniMax, StepFun, Baidu, iFLYTEK Spark, Meituan, AIsphere, and Shengshu Technology.
• Feishu (Lark)coreTechnology or Integration · 1 March 2026Feishu rapidly capitalized on the AI agent trend by launching an integrated 'Intelligent Partner' (aily) product alongside six connected AI model manufacturers including Kimi, positioning the platform as the operating layer for AI agents in organizational contexts.
• NVIDIAflagshipStrategic or Co-development Partner · 1 January 2026NVIDIA featured Kimi at GTC 2026, providing significant validation for the company. This featured placement positioned Kimi as a leading independent AI company alongside major industry players, helping drive the company's valuation growth.
• CloudflarecoreTechnology or Integration · 1 January 2026Cloudflare integrated Kimi K2.5 into their infrastructure and achieved 77% inference cost reduction when switching to K2.5, demonstrating the model's production-readiness and cost advantages.
• PerplexitycoreTechnology or Integration · 1 January 2026Perplexity validated Kimi as part of the company's remarkable turnaround strategy, using Kimi models as a key component of their AI search infrastructure.
• CursorcoreOEM/ Whitelabel/ Licensing Partner · 1 January 2026Cursor's Composer 2 coding model was built through fine-tuning on Kimi K2.5. The model was legally licensed through Fireworks AI. This partnership placed Kimi as the underlying engine for one of the most popular coding tools.

Scale indicators6 records

Recent moves6 records

Expansion highlights6 records

Others

• Moonshot AI / Kimi (self-reference handled at peer level):

Broad incumbents

• Anthropic: Closed-source frontier model developer (Claude) and primary reference point for Kimi's 'cost-effective Claude alternative' value proposition. Comparable on coding/agentic positioning and enterprise API customers.
• OpenAI: Frontier closed-source LLM leader (GPT-5 referenced in Kimi's Humanity's Last Exam comparison). Comparable on benchmark category leadership, API monetization, and enterprise inference spend — the closed-source incumbent K2.5 competes against on cost.
• Meta (Llama): Sets the global open-weight baseline (Llama family). Comparable on open-source distribution and developer ecosystem leverage, though backed by Meta's much larger distribution and compute footprint.

Direct peers

• Alibaba Qwen: Alibaba's open-source LLM family and a direct competitor in the Chinese open-weight frontier-model race. Comparable on multilingual capability, coding benchmarks, and open-source distribution; also an investor and customer of Kimi, creating overlapping commercial relationships.
• Zhipu AI: Beijing-based Chinese AI lab with competitive open-weight and enterprise offerings. Directly comparable on agentic capabilities, Chinese enterprise distribution, and Huawei Cloud / domestic cloud-channel partnerships.
• DeepSeek: Chinese open-weight LLM developer whose 2025 release cycle nearly drove Kimi out of the market. Directly comparable on open-source MoE architecture, Chinese developer base, and price-aggressive go-to-market targeting enterprise inference cost reduction.
• StepFun: Chinese multimodal AI startup with comparable open-weight releases and Huawei Cloud partnership. Directly competes in the domestic Chinese agentic/foundation-model category.

Emerging players

• xAI: Frontier model developer with comparable Grok model offering and aggressive scaling trajectory. Comparable on multimodal model portfolio and direct integration into consumer/developer surfaces.
• Mistral AI: European open-weight model developer with a comparable open-source + commercial API strategy. Comparable on open-weight release model, MoE architecture experimentation, and developer-led GTM.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights6 records

Customer concentration

• GitHubgithub.com/kimi-k2/Kimi-K2

Revenue estimate

Valuation estimate

Number of profiles

Funding overview

Funding rounds

Investors

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Kimi K2