AI Expertise
Open Models / Kimi K3
K

Kimi K3

Visit website

Description

Kimi K3 is an open language model from Moonshot AI, released on 2026-06-13. The card calls it the company's most capable model and the world's first open model of the 3-trillion-parameter class: it has about 2.8 trillion parameters. It understands text, images and video, and its context window is one million tokens. The weights are stored in a compact MXFP4 format that the authors trained the model for from the start.

The model is built for long programming sessions, work with large code repositories, terminal tools, knowledge work and hard reasoning. Thinking is always on; its depth can be set to low, high or maximum. In long conversations the full reply from the model, including its reasoning and tool calls, has to be passed back as it is. The authors recommend running it with vLLM, SGLang or TokenSpeed. The card does not state hardware requirements; with this size it is a server-level model.

The weights are public without any acceptance step. The Kimi K3 License lets you use, change, fine-tune and sell the model for free, with conditions: keep the license notice; if you offer the model to others through an API or similar and your company's revenue is above 20 million US dollars over 12 months, sign a separate agreement with Moonshot AI; and if a product built on it has more than 100 million monthly users or more than 20 million US dollars of monthly revenue, show the name "Kimi K3" in the product interface. Internal use and use through the developer's own products or certified partners are exempt from the last two rules. Because of these conditions we mark commercial use as limited. We found no territorial restrictions in the license text; read the license before use.

Date of check: 2026-10-21.

Pros and cons

Pros

  • One of the strongest open models: about 2.8 trillion parameters
  • Understands text, images and video; one million tokens of context
  • Depth of reasoning: low, high or maximum
  • Free use for most companies; the extra rules apply only above large thresholds
  • Recipes for vLLM, SGLang and TokenSpeed

Cons

  • Huge model: needs a serious server, not a home computer
  • The card does not state hardware requirements
  • License conditions: separate agreement for large API businesses, name display for very large products
  • Thinking cannot be switched off
  • Audio input is not mentioned in the card

Key facts

License
Kimi K3 License
Commercial use
Allowed with restrictions
Developer
Moonshot AI
Interface languages
English
Checked on
Oct 21, 2026

Editorial score

How strongly the AI Expertise editors recommend this open model for each of its tasks right now, on a 0–10 scale. The score combines output quality, capabilities, ease of launch, license openness and freshness.

Language models

6.9 / 10
Output quality 9.0
Low High
Features 8.6
Basic Rich
Ease of launch 3.5
Hard Easy
License openness 5.0
Restricted Open
Freshness 8.0
Outdated Fresh

News about Kimi K3

  • Chinese AI firm Moonshot AI launches Kimi K3.1 Pro, scoring 55 on the Artificial Analysis Intelligence Index — 4th globally, just 3 points behind Claude Opus 5.5. Massive 11-point jump over K3, driven by post-training compute. Leads in agentic tasks (GDPval-AA 1846 Elo), coding (Terminal-Bench 65.7%), 3D modeling (BenchCAD 92.7%), and office tasks. Same 2.8T params, 33% cheaper API ($2/M in, $10/M out). Weights planned for open release by Nov 2026 — would make it the top open-weights model, cutting the open-vs-prop gap from 12 to 3 pts. Cyber data stripped for safety. #MoonshotAI #KimiK3 #OpenWeights #AIbenchmark #LLM
  • China's Moonshot AI released Kimi K3, a model rivaling top Western systems despite US chip sanctions. With 2.8T parameters, it matches Anthropic's Opus 4.8 and challenges the "compute moat" thesis — scarcity drove architectural innovation (Mooncake stack, efficient RL). DeepMind researcher calls it "insanely good," not explainable by distillation alone. OpenAI strategist warns open-weight models could lead to "full AI communism," predicts US "soft law" regulatory friction instead of bans. Jevons paradox suggests efficiency may increase, not decrease, compute demand. Nobody knows what happens next. #AI #KimiK3 #ChinaAI #ComputeMoat #JevonsParadox
  • Moonshot pauses new Kimi K3 subscriptions after GPU capacity maxed out in 48 hours. Current users unaffected; new slots reopening gradually. Startup splits service into "Kimi Membership" (web/app/work) and "Kimi Code Membership" (coding) to balance compute load. Meanwhile, Alibaba launches Qwen 3.8 open-weight with discounted preview. Demand for AI inference still outpaces supply. #AI #KimiK3 #Moonshot #GPUShortage #Qwen3
All news about Kimi K3