What Was Released
<cite index="1-2">Moonshot AI, the Beijing-based artificial intelligence startup backed by Alibaba, on Thursday released Kimi K3 — a 2.8-trillion-parameter model that the company says is now the largest open-source AI model in the world, and one that benchmarks show performs neck-and-neck with the most powerful proprietary systems from Anthropic and OpenAI.</cite>
<cite index="22-4">The model relies on a Mixture-of-Experts (MoE) architecture, activating 16 out of 896 experts to manage compute resources efficiently.</cite> <cite index="23-4">Because only a fraction of its 2.8 trillion parameters activate on any given request, inference costs are reduced compared to a dense model of equivalent size.</cite> <cite index="1-7">The model features a 1-million-token context window, native visual understanding capabilities, and an always-on reasoning mode that the company calls "thinking mode."</cite>
Two internal architectural innovations underpin the design. <cite index="1-8">The model is built on Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which the company describes as a drop-in replacement for residual connections that delivers consistent scaling gains.</cite> <cite index="18-12">Moonshot reports a roughly 25% training-efficiency gain for under 2% additional compute overhead, which helped make training a 2.8-trillion-parameter model economically feasible under China's compute constraints.</cite>
Scale in Context
<cite index="1-6">Kimi K3 is a frontier-class large language model (LLM) with 2.8 trillion total parameters — roughly 75 percent larger than DeepSeek's V4 Pro, which the company's own timeline chart shows at approximately 1.6 trillion parameters.</cite> <cite index="9-1">The parameter count is significantly larger than previous open models from Chinese competitors, including DeepSeek's 1.6 trillion-parameter V4 Pro or Zhipu AI's 744 billion GLM 5 series, making it the largest open-source model so far.</cite>
<cite index="8-1">Moonshot AI publicly released Kimi K3 on July 16, 2026, with full open-source weights promised by July 27.</cite> Until then, the model is accessible exclusively via API. <cite index="1-10">On the API side, Kimi K3 is compatible with the OpenAI software development kit (SDK), lowering the integration barrier for developers already building on OpenAI or Anthropic toolchains.</cite>
Pricing and Access
<cite index="17-1">Kimi K3 is Moonshot's flagship model for long-horizon coding and end-to-end knowledge work, with a 1M-token context window and industry-leading intelligence, according to the company's official API documentation.</cite> <cite index="12-1">Pricing is set at $3.00 per million input tokens and $15.00 per million output through the Moonshot AI API, with a cache hit dropping input to $0.30 — a 90% discount — and the price is flat across the whole 1M-token context window.</cite>
<cite index="29-5,29-6">At $15 per million output tokens, K3 is considerably more expensive than z.ai's GLM-5.2 at $4.40 and DeepSeek V4 at $0.87, though still cheaper than Anthropic's Fable 5, which costs $50 per million output tokens.</cite> <cite index="18-15,18-16">Its predecessor, Kimi K2.6, ran roughly $0.95 input / $4 output per million tokens, so K3 represents a substantial step up that commentators have read as the end of "super-cheap Chinese AI" — Moonshot is now pricing on capability rather than undercutting.</cite>
Benchmark Performance
<cite index="9-3,9-4">K3 was designed for long-horizon coding, knowledge work, and reasoning tasks. While acknowledging that K3's "overall performance still trails the most powerful proprietary models," the company said the model had demonstrated frontier-level performance across a range of evaluations, "consistently outperforming other tested models," including OpenAI's GPT-5.5, Anthropic's Claude Opus 4.8, and rival Zhipu AI's latest GLM-5.2.</cite>
<cite index="16-6">Kimi K3 scores 57 on the Artificial Analysis Intelligence Index, placing it well above average among comparable models.</cite> <cite index="8-10">K3 leads all models on SWE Marathon and Program Bench, suggesting particular strength in sustained coding sessions — consistent with the 1M-token context window enabling full-repository understanding.</cite>
Company Background and Market Reaction
<cite index="25-27">Moonshot AI is a private Beijing company founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin.</cite> <cite index="23-3">Backed by Alibaba, Tencent, and Meituan, the company raised $2 billion at a $20 billion valuation in May and is now in talks for a round that would value the company at $30 billion.</cite> <cite index="29-8">A statement from the company's financial advisor stated Moonshot's annual recurring revenue (ARR) exceeded $200 million.</cite>
<cite index="10-16">Running a 2.8-trillion-parameter model locally could require hundreds of thousands of dollars in computing equipment, according to an estimate cited by Reuters, putting full on-premises deployment beyond the reach of many organizations.</cite> The July 27 open-weight release will determine whether independent evaluators can reproduce Moonshot's benchmark claims and whether the broader developer community can adapt the model for specialized use cases.
<cite index="1-3">The release was timed to land just ahead of the 2026 World Artificial Intelligence Conference in Shanghai.</cite> <cite index="26-4">Nvidia shares fell by 1.2% following the announcement, briefly forfeiting the chipmaker's lead as the world's most valuable company to Apple.</cite>