Models
Kimi K3: Why Millions of Developers Are Watching Moonshot AI’s Open Weights
Kimi is the model family from Chinese company Moonshot AI. In summer 2026, the company released Kimi K3, a 2.8-trillion-parameter model with a one-million-token context window and, more notably, open weights—the first open model in the “3 trillion” class. Here is what that means for everyday users and developers.
Current Kimi models
Kimi K3 launched on July 16, 2026; Moonshot published its weights and technical report on July 27. Its mixture-of-experts design has 896 experts, with 16 active at a time, plus a hybrid KDA (Kimi Delta Attention) mechanism. At launch, it was the world’s largest open model.
Prices rose from K2.6’s $0.95/$4 to $3/$15. Moonshot is targeting the Claude Sonnet segment rather than competing to be the cheapest model. K2.7 Code remains a lower-priced specialist for programming.
| Model | What it is | Price per 1M tokens |
|---|---|---|
| Kimi K3 | Flagship: 2.8T parameters (MoE), 1M context, open weights | $3 input / $15 output |
| Kimi K2.7 Code | Specialized in coding and agent tasks | Cheaper than the flagship; high-speed option available |
| Kimi K2.6 | Previous generation with multimodal input | $0.95 / $4 |
How capable is it?
Moonshot’s own benchmarks show K3 ahead of Claude Opus 4.8 and GPT-5.5, but behind Claude Fable 5 and GPT-5.6 Sol—placing it in the upper-middle tier, near the flagships. Artificial Analysis reports a similar picture: K3 ranks second to Claude Fable 5 on long-running tasks.
For developers, K3 reached the top of Arena.ai Frontend Code, a public ranking of web-interface generation. It is one of the stronger tools for turning a description into a working frontend. On agent tasks, it uses 21% fewer output tokens than its predecessor for the same amount of work.
Safety and refusals
Kimi tends to refuse less often than Western flagship models on controversial technical and creative topics, which is one reason for its popularity. Political restrictions associated with models trained in China still apply, though they are rarely noticeable in everyday tasks.
Open weights let you deploy the model on your own infrastructure and remove the service layer, which can help with privacy. You also take responsibility for use and quality control, while Moonshot’s built-in service filters are no longer present.
After K3 launched, demand was so high that Moonshot temporarily paused new subscription registrations because it lacked GPU capacity. Availability can fluctuate during peak periods.
What does it cost?
Artificial Analysis estimates an agent task at about $0.94 with K3, $1.04 with GPT-5.6 Sol, and $1.80 with Claude Opus 4.8. That puts K3 near OpenAI on task cost and about half the price of Claude at similar quality.
If you need to keep costs down, consider K2.7 Code for programming. This specialized model is considerably cheaper than the flagship and may be capable enough for many teams’ coding tasks.
Key features and limitations
- A 1,048,576-token (1M) context window can hold documents, repositories, and long conversations.
- It accepts text and images and produces text only.
- Open weights use the Kimi K3 license. Commercial use requires a separate agreement; review the terms before integrating it into a product.
- Generation speed is about 32 tokens per second. The first token can take a long time because the model reasons before responding, which is noticeable in interactive chat.
- There is one reasoning-depth level, without the flexible effort controls offered by OpenAI and Anthropic.
Who should consider Kimi?
Kimi K3 is an option if you want near-flagship capability at a moderate price for long analysis, frontend generation, or large Russian-language documents. Its language support is strong, including Russian conversations.
Developers may value open access: the weights have been downloaded a million times from Hugging Face, and major cloud providers have discussed K3-based services. The model is too large for most home computers and is better suited to a cluster. For local use, consider smaller Qwen models. For predictable behavior in strict corporate workflows, consider Claude, which can make more context-aware refusals. Kimi, Claude, and current OpenAI and Google models are in the Any LLM catalog; compare them with the same prompt in the chat.
Frequently asked questions
Is Kimi K3 free?
The weights are open to download, but running the model requires substantial hardware. The kimi.com site and API charge for the model: $3/$15 per million tokens, with limited free access on the site.
How is Kimi different from ChatGPT?
Kimi can cost less per task and tends to refuse less often; OpenAI offers a broader product range and more consistent service. The companies’ flagship models are close in raw capability.
Can I use K3 in a commercial product?
The weights use the Kimi K3 license, which requires a separate agreement with Moonshot for commercial use. Read the license before embedding the model in a product.
Why is Kimi slow to respond?
The model spends time reasoning before it answers. On complex tasks, time to first token can be measured in minutes. This is noticeable in interactive chat but less of a concern for background processing.
Which should I choose: Kimi K3 or Qwen3.8-Max?
Both are Chinese models with 1M context windows. K3 can cost less per task and has open weights; Qwen is stronger in Russian and offers a broader lineup. Try both on your own task in the Any LLM chat.