Moonshot AI launches 2.8 trillion parameter open-source Kimi K3 to challenge Anthropic

Kimi K3 is the largest open-weight AI model ever released, directly challenging Anthropic's closed-source Claude Opus 4.8.

Moonshot AI's Kimi K3, a 2.8 trillion-parameter open-weight model, matches or beats Anthropic's Claude Opus 4.8 on key benchmarks while costing a fraction to deploy, threatening the US lab's pricing power.

"Kimi K3 closes the gap between Chinese open-source models and the best closed-source systems to a degree that challenges the prevailing 8-to-12-month lag narrative," a person familiar with the company's internal positioning told the Financial Times.

The model uses a mixture-of-experts architecture with 2.8 trillion total parameters and a 1 million-token context window, compared with an estimated 1.5 trillion to 2 trillion for Opus 4.8. On SWE-bench Verified, LiveCodeBench, Tau2 for agent tasks, and AIME for math reasoning, Kimi K3 reached leading open-source levels and approached or matched top closed-source models, according to Moonshot AI's published results.

The timing amplifies the competitive pressure. Anthropic plans to raise Opus 4.8 prices by about 50% starting September, to $3 per million input tokens and $15 per million output tokens. Moonshot AI's earlier K2.6 model cost roughly one-third of Opus 4.8. If Kimi K3 follows a similar pricing curve, enterprises running large-scale inference workloads could see cost savings in the millions annually by switching.

The open-weight release — developers can download, deploy, and fine-tune the model freely — mirrors the strategy that propelled DeepSeek to global prominence. DeepSeek is now raising funds at a roughly $71 billion valuation, while Moonshot AI is in the process of raising capital at about $31.5 billion. By comparison, Anthropic's latest fundraising round valued it at $96.5 billion and OpenAI at $85.2 billion.

The Pricing Gap Widens

The cost differential is the most immediate threat to US AI labs' business models. Anthropic's planned 50% price increase for Opus 4.8 comes as Chinese competitors push costs lower. DeepSeek's R1 series already demonstrated that open-weight models can undercut US pricing while maintaining competitive capability on many tasks. Marc Andreessen, co-founder of Andreessen Horowitz, said Chinese AI lab Zhipu's GLM-5.2 had become the first Chinese model to match or lead US flagship models on several public benchmarks.

Kimi K3 has not yet published official benchmark results against Anthropic's unreleased Fable model, which was paused over safety concerns. According to industry estimates, Fable remains decisively ahead on raw capability. But Kimi K3 wins on price "by an order of magnitude" and on openness outright, according to one analysis. The model's 1 million-token context window — four times the size of GPT-4's standard context — makes it particularly suited for long-horizon software engineering tasks, a use case where enterprises have traditionally relied on closed-source models.

What This Means for Investors

For investors, the key question is whether US AI labs can maintain pricing power as open-weight alternatives improve. Nvidia, whose GPUs power most AI training and inference, benefits from increased model-building activity regardless of which company wins. But Anthropic and OpenAI face pressure to justify premium pricing when free, competitive alternatives exist. Moonshot AI's $31.5 billion valuation target and DeepSeek's $71 billion target suggest the market is already pricing in a shift toward open ecosystems that could reshape the AI industry's revenue model. The Financial Times reported that more Silicon Valley investors and tech executives now believe the real competitive battleground is developer ecosystems, not benchmark scores.

This article is for informational purposes only and does not constitute investment advice.