Beijing-based AI startup Moonshot AI, backed by Alibaba, has released Kimi K3, a language model with 2.8 trillion parameters that stands as the largest open-source model ever created. Benchmarks show it performs on par with the most powerful proprietary systems from Anthropic and OpenAI, marking a pivotal moment in the global AI race. The launch, timed just before the 2026 World AI Conference in Shanghai, is a dramatic comeback for a company that lost market ground after DeepSeek's rise.
Architecture with 2.8 Trillion Parameters and Million-Token Context Window
Kimi K3 boasts a cutting-edge architecture with 2.8 trillion total parameters, roughly 75% larger than DeepSeek's V4 Pro at 1.6 trillion. It features a million-token context window, native visual understanding, and an always-on reasoning mode called "thinking mode." Key innovations include Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, drop-in replacements for residual connections that improve scaling. Both techniques were published as open research on GitHub.
Sponsored Protocol
Top-Tier Benchmark Performance and OpenAI SDK Compatibility
Benchmark results place Kimi K3 third on GDPval-AA v2 with 1,687 points, behind only Claude Fable 5 Max and GPT-5.6 Sol Max. On AA-Briefcase, a long-horizon knowledge work benchmark, K3 scores second with 1,527 points, beating GPT-5.6 Sol Max. It achieves a record 91.2 out of 100 on BrowseComp, a complex information-seeking task, all in a single-agent setup without context compression. As one analyst noted: "Open source is no longer six months behind Western closed-source models." The model is compatible with the OpenAI SDK, easing developer integration. Pricing is $3 per million input tokens and $15 per million output tokens, with cached inputs at just $0.30.
Autonomous Chip Design Demonstration in 48 Hours
Moonshot AI showcased K3's capabilities with a proof-of-concept: the model independently designed a physical chip to run a nano-scale version of itself. Over 48 hours of continuous operation, it completed the full design pipeline using open-source EDA tools. The resulting 4-square-millimeter chip achieved timing convergence at 100 MHz and decoded over 8,700 tokens per second in simulation. This long-range autonomous agent ability marks a leap beyond single-turn question answering. In another case, K3 reproduced the universal I-Love-Q relation in computational astrophysics in about two hours, reading and cross-validating over 20 papers.
Sponsored Protocol
Moonshot AI's Comeback After DeepSeek's Disruption
Founded in 2023 by Yang Zhilin, a former researcher at Google and Meta, Moonshot AI quickly became a leading Chinese AI startup, raising about $1.5 billion by early 2026. However, DeepSeek's R1 model in January 2025 disrupted the market, dropping Kimi from third to seventh in monthly active users. The company pivoted to open-source, releasing Kimi K2 in July 2025 and K2.5 in January 2026. Kimi K3 is the culmination of this strategy. As covered in an earlier article, China's AI acceleration is reshaping global dynamics.
Sponsored Protocol
Geopolitical Implications and Enterprise AI Future
The decision to release full weights on July 27 is strategic. Kimi K3 offers enterprises an open-source alternative to Western APIs, though it requires substantial GPU infrastructure. The Mooncake project, awarded Best Paper at FAST 2025, improves inference efficiency. With a model lineup including K3 (flagship), K2.7 Code (coding), and K2.6 (general), Moonshot targets developers and enterprise contracts. The gap between open-source and proprietary models has effectively closed, and algorithmic innovation compensates for chip export restrictions. For more on large language models, see Wikipedia.
