Moonshot AI's Kimi K3 debuts with 2.8 trillion parameters, challenging Anthropic and OpenAI

China's Moonshot AI has released Kimi K3, a massive mixture-of-experts model with 2.8 trillion total parameters and a 1 million-token context window — placing it directly in competition with Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol for dominance in the frontier AI market.

A model built for scale

Kimi K3 uses a sparse MoE architecture with 896 experts, but only 16 are activated for any given token — roughly 1.8% of the pool at any moment. This means the model can hold an enormous knowledge base while keeping inference costs manageable. The result is competitive performance at significantly lower per-token cost than comparable dense models.

Two variants launched simultaneously: K3 Max, optimized for conversational and agentic tasks, and K3 Swarm Max, designed for large-scale parallel processing workflows. Both accept text, image, and video as input.

Where it stands on benchmarks

Within days of launch, Kimi K3 debuted at number three on the Artificial Analysis AI leaderboard, trailing only Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol. It pulled ahead of both on Arena.ai's front-end web development benchmark, suggesting particular strength in code generation tasks.

API pricing is set at $3 per million input tokens and $15 per million output tokens — undercutting the premium tier pricing of both Anthropic and OpenAI's flagship models at equivalent capability levels.

Open weights coming July 27

Moonshot AI has committed to releasing open weights for Kimi K3 by July 27, 2026 — a move that would make it by far the largest openly available Chinese AI model, and potentially the largest open-weight model globally. That release would allow developers to run and fine-tune the model without relying on Moonshot's API.

The company is currently fundraising at a $31.5 billion valuation, up from the $20 billion valuation at which it raised $2 billion in May 2026. The rapid cadence of model releases — and a model that can legitimately challenge top US labs — is clearly driving investor appetite.

Context: China's AI gap is narrowing fast

Kimi K3 follows a string of competitive Chinese AI releases over the past year, including DeepSeek's cost-efficient models and Alibaba's Qwen series. What distinguishes K3 is its scale: 2.8 trillion parameters puts it in the same weight class as the largest US models, not just a competent second tier.

US officials have pushed to restrict Chinese access to advanced AI chips, but Moonshot's Kimi K3 suggests that workarounds — whether through stockpiled hardware, alternative supply chains, or architectural efficiency — are allowing Chinese labs to compete at the frontier regardless. The open-weight release on July 27 will reveal how the model performs under independent evaluation outside of Moonshot's controlled benchmarks.