Moonshot AI publicly launched Kimi K3 on July 16, 2026, a 2.8 trillion parameter open-weight AI model designed for code generation and multimodal tasks [1, 2, 3]. The model features a 1 million token context window and supports both text and image inputs natively [1, 2].
Kimi K3 tops Arena’s frontend coding leaderboard, surpassing leading closed-source models including Anthropic’s Fable 5 and OpenAI’s GPT-5.6 Sol, marking a milestone in open-weight AI capabilities [2, 3]. It is built on new architectural techniques, including Kimi Delta Attention, Attention Residuals, and scaled-up Mixture of Experts sparsity, boosting efficiency and performance [1]. Moonshot reports a 2.5-fold scaling efficiency gain over its predecessor, Kimi K2 [1].
Pricing for Kimi K3 is set around $12 per million tokens, aligning more closely with mid-tier Anthropic models rather than the steeper discounts Moonshot offered before [2]. Early benchmarks and demo results have been promising but require further real-world testing, as the full model weights were not released at launch [2, 3].
Moonshot AI is collaborating with inference partners and open-source maintainers to ensure a stable rollout and wider accessibility when the weights become available. The full model weights are scheduled for public release on July 27, 2026 [1, 2, 3].
Mozilla CTO Raffi Krikorian noted the competitive implications, saying, "Right now, it's a U.S. versus China question. U.S. AI labs are clearly worried" about Moonshot’s progress [2].
The next key milestone for the model’s wider adoption is the planned release of Kimi K3’s full weights later this month, which will allow the AI community to begin extensive real-world testing and deployment [1, 2, 3].
