Moonshot AI Drops Kimi K3: World's Largest Open-Source AI Model With 2.8 Trillion Parameters

Posted on 18th Jul 2026 09:43:27 in Artificial Intelligence, Machine Learning

Tagged as: AI, Moonshot AI, Kimi K3, open source, LLM, China AI, frontier model

China's AI ecosystem just delivered its loudest statement yet. Beijing-based Moonshot AI, backed by Alibaba, released Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter Mixture-of-Experts model that the company calls the largest open-source AI system ever built. The release, timed just hours before the World Artificial Intelligence Conference in Shanghai, sent shockwaves through global markets and ignited a fresh wave of debate about the narrowing gap between Chinese and American frontier AI labs.

K3 is not just bigger than its competitors. It is 75 percent larger than DeepSeek's V4 Pro at roughly 1.6 trillion parameters, nearly triple Moonshot's own Kimi K2.6, and well past Zhipu AI's 744-billion-parameter GLM 5 series. The model ships with a 1-million-token context window, native visual understanding, and two architectural innovations developed in-house: Kimi Delta Attention and Attention Residuals. Full model weights are scheduled for release on July 27.

Architecture and Technical Innovations

Kimi K3 is built on a Mixture-of-Experts architecture with 896 expert networks, of which only 16 are activated for any given inference request. This sparse activation pattern means the model can route individual tasks to specialized sub-networks without engaging its full parameter count on every query.

The two architectural breakthroughs merit closer attention. Kimi Delta Attention is a hybrid linear attention mechanism that delivers up to 6.3 times faster decoding in million-token contexts compared to conventional attention implementations. Attention Residuals, the second innovation, is a drop-in replacement for standard residual connections in transformer blocks, delivering approximately 25 percent higher training efficiency at less than 2 percent additional computational overhead.

The model features an always-on reasoning mode called "thinking mode" and supports the OpenAI-compatible API format. Early testers reported inference speeds around 28 tokens per second with multi-second first-token latency.

Benchmark Performance: Trading Blows With the Frontier

On the GDPval-AA composite benchmark, Kimi K3 scored 1,687 — third overall behind only Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. Independent testing from Artificial Analysis placed K3's intelligence score behind only Fable 5 on its Intelligence Index.

Coding is where K3 truly shines. The model claimed first place on ProgramBench and SWE-Marathon, two benchmarks evaluating an AI's ability to construct complete software projects. On Terminal-Bench 2.1, K3 trailed GPT-5.6 Sol by only half a point. On the Fronted Code Arena, K3 jumped from 18th place for its predecessor to first overall, topping six of seven measured domains.

Moonshot demonstrated the model completing multi-step tasks including building playable browser-based 3D games, a functional Game Boy Advance emulator, multiplayer arena shooters, and even designing a functional computer chip through an autonomous multi-hour run.

Pricing and Strategic Positioning

At $3 per million input tokens and $15 per million output tokens with cached input at $0.30 per million, K3 is the highest-priced Chinese AI model yet — but still substantially undercuts Western equivalents. Anthropic's Claude Opus 4.8 costs $5/$25, and OpenAI's GPT-5.5 sits at $5/$30. This positions K3 as a frontier-tier flagship that happens to be cheaper.

According to Artificial Analysis's Cost per Intelligence metric, K3 averages about $0.94 per equivalent benchmark task — comparable to GPT-5.6 Sol and roughly half the cost of Claude Opus 4.8. Developers can access K3 immediately through the Kimi web app, the official API using an OpenAI-compatible endpoint, or through OpenRouter.

Geopolitical Context and Market Impact

The K3 release landed in a highly charged environment. The World AI Conference opened in Shanghai on July 17 with Chinese President Xi Jinping delivering his first-ever keynote at the event. The market reaction was immediate — news of K3's capabilities triggered a sell-off in US technology stocks, with Bloomberg reporting the "surprise breakthrough from a Chinese AI startup rippled through global markets."

The episode echoed the DeepSeek shock of early 2025, but this time the threat is more substantive: K3 is open-weight and demonstrably competitive at the frontier. Once the open weights land on July 27, any team with sufficient hardware can run frontier-adjacent AI without sending code or data to a third-party API.

What Comes Next

The open weights drop on July 27 will enable self-hosting, fine-tuning, and widespread independent benchmarking. The WAIC continues through July 20 with expectations of further announcements. Longer-term, the absence of a published CyberGym score raises questions the global AI policy community will need to address about safety evaluation standards for open-weight frontier models.

Sources

• VentureBeat — China's Moonshot AI releases Kimi K3, the largest open-source model ever (July 16, 2026)
• South China Morning Post — Moonshot AI unveils world's largest open-source AI model (July 16, 2026)
• The Indian Express — Moonshot AI unveils Kimi K3, the world's largest open-weight AI model (July 17, 2026)
• In The World of AI — Kimi K3 Drops: 2.8T Open Model Rivals GPT-5 (July 16, 2026)
• Bloomberg — China's New AI Fuels Tech Rout (July 17, 2026)