AI Knowledge · Kimi Guide
Kimi K3 by Moonshot AI:
The Business Guide
Kimi K3 is the first open model from China to break into the top tier of AI systems. This guide answers the questions founders and teams are asking now: how good is the model really, what does it cost, what does the China location mean for data protection, and when is it worth looking beyond Claude and ChatGPT. Benchmarks and prices on this page update automatically every day.
What is Kimi K3?
Kimi K3 is the flagship model of Moonshot AI, an AI company from Beijing founded in 2023 and backed among others by Alibaba. In China, the Kimi chatbot has been one of the most used AI assistants for years. Internationally, Moonshot built its reputation with the open K series, whose model weights are freely available. On July 16, 2026, Moonshot introduced Kimi K3. The technical cornerstones:
- Mixture-of-experts architecture with roughly 2.8 trillion parameters, the largest open model to date. Only 16 of its 896 experts compute per token, which works out to around 104 billion active parameters.
- A one-million-token context window: entire contract sets, reports, or codebases in a single pass.
- An always-on thinking mode: the model reasons before every answer.
- Multimodal: Kimi K3 natively understands text, images, and video.
- Open weights: Moonshot published the model weights on July 27, 2026 (1.56 terabytes in the compact MXFP4 format), so the model can now be operated outside China.
For companies, that last property is the most interesting one: a model at the level of the Western frontier class that can in principle run on your own or European infrastructure changes your negotiating position with every AI vendor.
How good is Kimi K3? Today's numbers
Vendor claims age quickly, so the numbers on this page do not come from a press release. We pull them daily from the independent measurements by Artificial Analysis and convert prices to euros using the ECB reference rate. In the current field, Kimi K3 ranks number 8 among the frontier models we track every day.
| Model | Vendor | Intelligence index | Price per 1M tokens | Open weights |
|---|---|---|---|---|
| Kimi K3 | Kimi | 60 | €5.14 | Yes |
| Claude Opus 5 | Anthropic | 63 | €8.57 | No |
| GPT-5.6 Sol | OpenAI | 61 | €6.86 | No |
As of August 27, 2026 · Source: Artificial Analysis (intelligence index, blended input/output price) and the ECB reference rate. This table updates automatically every day.
How Kimi K3 stacks up against every other system, sorted by use case, budget, and data protection, is shown in our daily updated AI Leaderboard; the methodology behind it is documented here.
Data protection and GDPR: the China factor
This is where testing and production part ways. If you use Kimi K3 via kimi.com or the Moonshot API, your data is processed in China, a country without an EU adequacy decision. That is fine for low-stakes experiments; for personal data, customer data, or trade secrets, this path is hard to justify under GDPR.
The open weights flip the picture: as soon as Kimi K3 runs with a European hosting provider or on your own infrastructure, your data never leaves the EU and Moonshot sees not a single prompt. The China risk turns into a sovereignty advantage that closed models cannot offer. Your homework stays the same as with every AI tool: classify data categories, extend your records of processing activities, and give the team clear rules about which data may go into which system.
Open weights and self-hosting: what is realistic
Moonshot released the Kimi K3 weights on July 27, 2026, free for anyone to download. Reality for mid-sized companies is more nuanced than the headline suggests: the published files take up 1.56 terabytes, and even the most aggressively compressed community build still asks for around 610 gigabytes of combined system and GPU memory. Running the full model means a server with eight top-end GPUs or roughly twenty older cards spread across several machines. That is inference-provider infrastructure, not a workstation under someone's desk.
- European inference providers: the most practical route, and no longer hypothetical. Within days of the release, Nebius listed the model on its Token Factory, and specialised providers run it via vLLM on eight B300-class GPUs in EU data centers, at roughly 0.40 euros per million input tokens and 0.60 euros per million output tokens, with a data processing agreement and a public list of subprocessors.
- Quantized variants: community builds now span from 594 gigabytes at about 79 percent of the original quality up to a lossless 1,560 gigabytes, so the trade-off is yours to choose. The smaller ones currently need a specific llama.cpp branch.
- The smaller siblings: Kimi K2.6 is the one-trillion-parameter allrounder, and Kimi K2.7 Code has been available since June 2026 for programming work, scoring 21.8 percent above K2.6 on Moonshot's own code benchmark while spending around 30 percent fewer reasoning tokens.
The open-weights label leaves one question open, and it is the one your legal team will ask: what does the license actually permit? Moonshot shipped the weights under its own Kimi K3 license instead of MIT or Apache 2.0, which is why Artificial Analysis lists the model as commercial use restricted. For a company running Kimi K3 on its own processes, documents, and automations, that label overstates the problem: purely internal use is explicitly exempt from every additional obligation. The thresholds target businesses that resell inference, and they start at 20 million US dollars of combined licensee and affiliate revenue over twelve consecutive months.
The weights are out, and so is our hands-on guide (in German): Self-hosting Kimi K3 after the open-weights release, covering the hardware math, European provider options, the license thresholds, and the GDPR assessment.
Frequently asked questions about Kimi K3
What is Kimi K3?
Kimi K3 is the flagship model that the Chinese AI company Moonshot AI introduced in July 2026. It is a mixture-of-experts model with roughly 2.8 trillion parameters, a context window of one million tokens, and an always-on thinking mode. The remarkable part: Moonshot publishes the model weights openly, which makes Kimi K3 the first open model that plays at the level of the leading Western frontier models in independent benchmarks.
How good is Kimi K3 compared to Claude and ChatGPT?
In independent measurements by Artificial Analysis, Kimi K3 sits close behind the best models from Anthropic and OpenAI and beats many established closed models. The exact numbers change constantly; the table on this page and our AI Leaderboard show the current standing, both updated automatically every day.
What does Kimi K3 cost?
The official Moonshot API bills by the amount of text processed (tokens), at prices clearly below comparably strong Western models. Since API prices change frequently, the table on this page shows the daily updated blended price. Since the weights release on July 27, 2026, a second cost logic applies: your own infrastructure or a European hosting provider instead of API billing, currently around 0.40 euros per million input tokens and 0.60 euros per million output tokens at EU providers.
Can Kimi K3 be used in a GDPR-compliant way?
When you use kimi.com or the Moonshot API, your data is processed in China, a country without an EU adequacy decision. That path is hard to justify for personal data or confidential business information. The realistic GDPR route runs through the open weights: run Kimi K3 with a European provider or on your own infrastructure, and your data never leaves the EU. Run that assessment before any production rollout.
Can I self-host Kimi K3?
Technically yes, in practice only with a multi-GPU cluster. The weights Moonshot published on July 27, 2026 take up 1.56 terabytes, because the model ships in the compact MXFP4 format rather than being compressed after training. The smallest community-quantized build comes down to 594 gigabytes and still asks for roughly 610 gigabytes of combined system and GPU memory. Realistic routes are European inference providers, quantized variants, or the smaller open sibling models Kimi K2.6 and Kimi K2.7 Code. Our hands-on guide with the full hardware math is linked on this page.
Is Kimi K3 free for commercial use?
For normal in-house use yes, without extra conditions. Moonshot released the weights on July 27, 2026 under its own Kimi K3 license, which reads like the MIT license across most of its text. Two special rules start far above the mid-market: anyone reselling model inference as a Model as a Service business needs a separate agreement once the combined revenue of the licensee and its affiliates passes 20 million US dollars over twelve consecutive months, and above 100 million monthly active users or 20 million US dollars in monthly revenue the name Kimi K3 has to appear visibly in the interface. Purely internal use is explicitly exempt from both. The smaller siblings K2.6 and K2.7 Code stay under the freer Modified MIT license and do not carry that threshold.
What is Moonshot AI?
Moonshot AI is an AI company from Beijing founded in 2023, backed among others by Alibaba. It became famous for the Kimi chatbot, one of the most used AI assistants in China. With its open K series (K2, K2.6, the coding model K2.7 Code, and now K3), Moonshot has built an international reputation and, alongside DeepSeek, is considered the most important driver of open frontier models from China.
Deep dives in our magazine
Our editorial team tracks the model market continuously. The most relevant pieces (in German):
- Kimi Skills: the curated catalog for Kimi Code (English)
- AI Leaderboard: the daily updated AI comparison (English)
- Self-hosting Kimi K3: hardware math, EU providers, license (German)
- Multi-provider strategy: why a second model is mandatory (German)
- Setting up AI model routing (German)
- Token costs in your company (German)
- Claude vs. ChatGPT for business use (German)
The right model strategy, without detours
We run AI models in production every day and measure daily which system leads where. If you want to know whether Kimi K3, Claude, or a combination is the right choice for your company, we build the system for you: model choice, data protection setup, routing, and a team that can work with it. Get in touch.
Let's talk