Kimi K3's 2.8 trillion parameters reshape AI cost dynamics

Moonshot AI's Kimi K3 delivers 2.8 trillion parameters at a fraction of the cost of US frontier models, reshaping the economics of AI.

Kimi K3 scores 91.2 on the BrowseComp benchmark — above Claude Fable 5's 88.0 and GPT-5.6 Sol's 90.4 — while costing $4.65 per rollout versus $13.41 for Anthropic's model, threatening the pricing power of US frontier labs.

"Open-weight models are closing the capability gap while opening the cost gap," said Jensen Huang, chief executive officer of Nvidia, in an interview with Axios. "US companies absolutely should be allowed to use Chinese models."

The model uses 896 routing experts in a mixture-of-experts architecture, activating 16 per token to keep compute efficient. Its Kimi Delta Attention mechanism replaces three-quarters of traditional attention layers with linear attention, achieving a 6.3x decoding speedup on million-token contexts. On the Kimi Code Bench, K3 scores 4 points below Claude Fable 5 but at 38% of its cost — delivering 2.8x more solved tasks per dollar. At a full 452-rollout sweep, K3 costs $2,103 versus $6,010 for Fable 5.

The release pressures US AI companies already trading at elevated multiples. Anthropic responded by releasing Claude Opus 5, a cheaper alternative to Fable 5. Nvidia formed a security alliance to promote safe open-weight deployment. Chinese models now account for about 60% of token usage by US companies on the OpenRouter API marketplace, according to platform data.

Architecture That Scales to 2.8 Trillion Parameters

The 93-layer network combines 69 KDA linear attention layers with 24 MLA layers in a 3-to-1 pattern. KDA maintains a fixed-size state buffer that updates with each new token, using a channel-level forget gate that lets different memory dimensions decay at different rates — entity relationships persist while local phrasing fades. A delta rule writes new information by comparing existing memory against the new value and correcting only the error. The system enforces a minimum per-step retention rate of about 0.67 percent, preventing numerical instability during parallel computation.

An Attention Residuals mechanism lets each layer retrieve outputs from any earlier layer across nine compressed blocks, solving the feature dilution problem that standard residual connections create at scale. In traditional 93-layer networks, roughly 44 percent of layers may produce minimal effective transformation, according to published research. AttnRes turns layer-to-layer information transfer from a fixed additive channel into a learnable routing system.

The Stable LatentMoE compresses token representations from 7,168 dimensions to 3,584 before routing to experts, keeping communication bandwidth flat despite doubling the active expert count from eight to 16. Two shared experts still operate at full width, processing every token.

US Adoption Grows Amid Political Backlash

DoorDash uses Kimi for lower-level tasks while reserving Anthropic's Fable for higher-level functions. Coinbase confirmed internal use. Cursor, the coding startup acquired by SpaceX, built its product on a Kimi model. The Trump administration responded negatively — science advisor Michael Kratsios accused Moonshot AI of illegally distilling Anthropic's Fable, a claim experts have disputed, while Treasury Secretary Scott Bessent threatened to add the company to the Entity List.

Moonshot AI closed a $3.5 billion funding round, according to reports, signaling investor confidence in the open-weight approach. The company's technical report provides a full recipe for training at the 3-trillion-parameter scale, including the MoonEP communication library that guarantees balanced expert allocation without training interruptions and a unified activation manager that dynamically decides whether to keep intermediate results on GPU, move them to CPU memory, compress to FP8, or recompute during backpropagation.

For investors, the math is straightforward. If K3 maintains a 2.8x cost advantage on coding tasks while matching or exceeding closed models on knowledge benchmarks, the addressable market for proprietary inference APIs shrinks. Nvidia, whose hardware powers both camps, remains insulated. But US AI companies trading at 30x-plus forward revenue multiples face a structural pricing headwind that no export control can eliminate.

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