Overview
<cite index="3-2">Moonshot AI, the Beijing-based artificial intelligence (AI) startup backed by Alibaba, released Kimi K3 on July 16, 2026—a 2.8-trillion-parameter model that the company says is now the largest open-source AI model in the world.</cite> <cite index="6-1">Moonshot describes Kimi K3 in its technical blog as the world's first open 3T-class system and the largest open-weight AI model to date.</cite>
Architecture and Specifications
Kimi K3 is built on a sparse Mixture-of-Experts (MoE) architecture. <cite index="6-4">The model has a 1-million-token context window, native vision, and activates just 16 of its 896 experts per token—roughly 1.8% of the pool.</cite> <cite index="2-5">The model introduces two architectural innovations: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), which Moonshot says improve efficiency and reasoning quality.</cite>
The weight file size reflects deliberate engineering choices. <cite index="12-4">Moonshot trained K3 to work at four bits of precision per parameter instead of the usual sixteen, a method known as quantization-aware training, which the company says it chose "for broad hardware compatibility."</cite> <cite index="12-6">Independent analysis of the release puts the model at roughly 1.4 terabytes (TB) in that format, against the 5.6 TB it would need at full precision.</cite> <cite index="5-7">Self-hosting the model would bring it within reach of organizations with multi-node Graphics Processing Unit (GPU) clusters, such as 8–16 nodes of 8x H100 or B200 accelerators.</cite>
Scale and Competitive Context
<cite index="2-4">Kimi K3 is roughly 2.8 times the size of its predecessor, K2.6, and dwarfs Chinese competitors including DeepSeek's V4 Pro at 1.6 trillion parameters and Zhipu AI's GLM 5 series at 744 billion parameters.</cite> <cite index="3-3">The release was timed to land just ahead of the 2026 World Artificial Intelligence Conference in Shanghai.</cite>
Benchmark Performance
<cite index="6-3">Moonshot said K3 still sits behind Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol on overall performance, but it outperformed every other model in the company's evaluation suite, including Claude Opus 4.8 and GPT 5.5, across coding and agentic benchmarks.</cite> <cite index="7-7">Kimi K3 took the number one spot in the Frontend Code Arena with 1,679 points, passing Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618.</cite> <cite index="11-3">On GDPval-AA v2, Kimi K3 scored 1,687, ranking third behind Claude Fable 5 Max and GPT-5.6 Sol Max, but ahead of Claude Opus 4.8.</cite>
Licensing, Pricing, and Access
<cite index="8-13">Moonshot committed to releasing the full model weights under a Modified MIT license by July 27, 2026.</cite> <cite index="1-4">Application programming interface (API) pricing is set at $3 per million input tokens and $15 per million output tokens.</cite> <cite index="20-3">Demand surged so quickly that Moonshot temporarily suspended new subscriptions after its computing capacity was overwhelmed.</cite>
Geopolitical Controversy and US Response
<cite index="18-3">Michael Kratsios, Director of the White House Office of Science and Technology Policy, stated on social media that Moonshot acquired servers equipped with Nvidia GB300 chips and routed workloads through GB300-enabled infrastructure in Thailand to train its models.</cite> <cite index="18-5">The semiconductor systems in question belong to Nvidia's Blackwell generation, specifically the GB300 platform, which the US government explicitly bars from export to Chinese companies.</cite> Kratsios also alleged that Moonshot distilled capabilities from Anthropic's Fable model at scale. <cite index="19-13">Moonshot has denied suggestions it used American models to train Kimi K3.</cite> <cite index="22-16,22-17">Several analysts have also expressed doubt about the accusation, arguing that the release schedule gave Moonshot limited time to collect enough data from newer US models before K3 entered the market.</cite>
<cite index="19-3,19-4">More broadly, Kimi K3's capabilities are renewing questions about the effectiveness of Washington's tech export curbs, and the new model challenges an industry belief that leading-edge AI requires heavy spending on data centers and the advanced chips that the US has restricted access to.</cite>
Company Background
<cite index="21-3,21-4">Moonshot AI was founded in 2023 and is among China's leading Large Language Model (LLM) developers, with backing from Chinese tech giants Alibaba and Tencent. The firm raised approximately $2 billion in its latest funding round, valuing it at more than $20 billion, according to Bloomberg.</cite> <cite index="21-6">Moonshot reported in April that its Annual Recurring Revenue (ARR) topped $200 million, driven by growing subscriptions for Kimi and its other AI services.</cite>