Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model With Kimi Delta Attention and 1M Context
Moonshot AI announced the Kimi K3 model on July 16, 2026. The open mixture‑of‑experts model has 2.8 trillion parameters, uses Kimi Delta Attention, and supports up to 1 million token context.
- Moonshot AI introduced Kimi K3, a 2.8 trillion‑parameter open MoE model.
- The model uses Kimi Delta Attention and supports a 1 million token context window.
- Kimi K3 activates 16 of 896 expert modules per token, aiming for efficiency at scale.
On July 16, 2026 Moonshot AI released Kimi K3, an open‑source mixture‑of‑experts (MoE) language model. The architecture scales to 2.8 trillion parameters and activates 16 of 896 expert modules per token.
Kimi K3 incorporates the proprietary Kimi Delta Attention mechanism and Attention Residuals, aiming to improve efficiency and long‑range coherence. It also expands the context window to one million tokens, far beyond typical 8‑to‑32 k token limits.
The model is publicly available for research and commercial use, positioning Moonshot AI as a competitor to other large‑scale open models. Early benchmarks suggest competitive performance on standard language tasks while offering lower inference cost per token due to the MoE design.
Provides a massive open model with long context for advanced applications.
Enables new products that require extensive document processing or long‑form generation.
Signals Moonshot AI's ambition to compete in the high‑parameter open‑model market.
Offers a research‑grade model for studying MoE architectures and attention variants.
Expands access to cutting‑edge language capabilities beyond proprietary services.
- MoE
- Mixture‑of‑Experts, a neural architecture that routes inputs to a subset of specialized sub‑models.
- Kimi Delta Attention
- A custom attention variant introduced by Moonshot AI to improve efficiency and long‑range dependencies.
China: Moonshot Release 2.8tn Parameter AI Rivalling the US - AI Magazine
Moonshot’s Kimi K3 closes the frontier gap - The Rundown AI
Chinese startup Moonshot AI unveils Kimi model it says rivals OpenAI, Anthropic - CNBC
LLMNVIDIA AI Releases Nemotron 3 Embed: An Open Embedding Collection Whose 8B Checkpoint Ranks #1 on RTEB
China's Moonshot unveils world's largest open AI model, closing in on US rivals - Yahoo Finance
AI giant Anthropic bringing new artificial intelligence for teachers to Detroit classrooms - WDET 101.9 FM
Anthropic, an AI giant, is introducing new AI tools for teachers in Detroit classrooms, aiming to enhance education.
Patreon stops asking AI bots not to scrape — and starts blocking them
Patreon is partnering with Cloudflare to block AI bots that scrape creators' content without permission, marking a shift away from relying on robots.txt.
HardwareNVIDIA Vera Rubin Maximizes Intelligence per Dollar for Post-Training Workloads – a Key Metric for Agentic AI
NVIDIA's Vera Rubin architecture aims to optimize intelligence per dollar, specifically targeting the high computational demands of post-training and agentic AI workflows.
CT redoubles efforts to equip workforce for AI economy - CT Mirror
The state of Connecticut is investing in programs to prepare its workforce for the AI economy.
Franz Inc. Named a 2026 Artificial Intelligence 100 Company by KMWorld - EIN News
Franz Inc. has been named a 2026 Artificial Intelligence 100 Company by KMWorld, a prestigious recognition in the AI industry.
Deep Reinforcement Learning-based combat recognition of traditional Chinese Sanda under artificial intelligence technology - Nature
Researchers from Nature used deep reinforcement learning to develop an AI system that can recognize traditional Chinese Sanda combat moves. The AI model was trained on a dataset of videos.