Kimi K3 Packs 2.8 Trillion Parameters, Challenging Closed US AI Models

Moonshot AI launched Kimi K3 on July 17, a 2.8-trillion-parameter open-weight model the Chinese startup said is the largest of its kind ever released. The model targets long-horizon coding, knowledge work, and reasoning tasks.

Benchmarks place its performance near Anthropic‘s frontier systems. The release is the most direct Chinese challenge yet to closed US lab models.

Kimi K3 Redraws the Open-Weight Frontier

Kimi K3’s 2.8-trillion-parameter scale matters because parameters are one proxy for a model’s raw capacity to store and apply learned patterns.

In a neural network, parameters are the numerical weights adjusted during training that determine how the model responds to any given input. More parameters generally allow a model to encode finer distinctions across language, code, and reasoning — though raw count is not the only factor in real-world performance.

Most open-weight models released to date have topped out well below one trillion parameters. By publishing weights at this scale, Moonshot makes the model available for outside researchers and companies to download, inspect, and run, unlike closed systems from OpenAI or Anthropic that are only accessible through paid APIs.

The architectural approach Moonshot used to reach this scale almost certainly involves a Mixture-of-Experts design, a technique in which only a subset of the model’s total parameters are activated for any given input.

This means the full 2.8 trillion parameters are not all running simultaneously, which would be computationally prohibitive. Instead, a routing mechanism selects the most relevant expert subnetworks for each token, allowing the model to maintain a vast learned knowledge base while keeping inference costs manageable.

DeepSeek’s most capable models use the same approach, which partly explains how Chinese labs have achieved frontier-scale results without matching the raw hardware budgets of US counterparts.

The model supports a 1-million-token context window, meaning it can process roughly 750,000 words of input in a single pass. That figure is significant for enterprise use cases such as legal document review, long-form code repositories, and extended scientific literature analysis.

A context window of this size allows the model to hold an entire large codebase or a multi-hundred-page document in active memory during a single query, rather than breaking the task into chunks that lose continuity across sessions.

Moonshot said Kimi K3 performs in English and Chinese, with benchmark scores it said approach Anthropic’s best available systems. No independent third-party evaluation of those results had been published as of Friday morning.

How China Closed the Open-Weight Gap

The open-weight AI race accelerated sharply after Meta’s Llama releases normalized the practice of publishing model weights publicly.

Chinese labs moved fast. DeepSeek drew global attention earlier this year when its R1 model matched top US systems on several coding benchmarks at a fraction of the reported training cost. That result demonstrated that algorithmic efficiency and architectural choices could compensate meaningfully for restricted access to the most advanced chips.

Kimi K3 is a direct escalation of that trajectory, adding scale that no prior open-weight release had reached.

Moonshot AI is a Beijing-based startup founded in 2023. The company’s earlier Kimi models built a consumer following in China through a long-context chat product before the company pivoted to frontier-scale research.

Kimi K3 is its first model positioned explicitly as a global research-tier release rather than a domestic consumer product.

Also Read: Grok 4.5 Arrives As xAI’s New Flagship For Coding And Agentic Work

What a 2.8-Trillion-Parameter Open Model Means for the AI Industry

The strategic consequence is distribution. Closed frontier models require users to route queries through a company’s servers, giving the lab visibility into usage patterns and a recurring revenue stream.

Open-weight models at Kimi K3’s scale shift that dynamic entirely. Any organization with sufficient compute can run the model internally, which is particularly attractive for governments, defense agencies, and enterprises with data-sovereignty requirements that prohibit sending sensitive information to third-party cloud infrastructure.

For US labs, the more immediate competitive pressure is on pricing and access.

When a credible open model matches a closed frontier system’s benchmark scores, the closed system’s price premium becomes harder to justify. DeepSeek’s release earlier this year already compressed API pricing across the industry.

Kimi K3’s launch at this parameter count repeats that dynamic at a larger scale.

The model has not yet undergone external safety evaluation. Anthropic and OpenAI both maintain internal red-teaming processes and publish safety reports before major releases.

Red-teaming involves structured adversarial testing by internal or contracted researchers who attempt to elicit harmful, deceptive, or dangerous outputs before a model reaches the public. Moonshot has not indicated whether an equivalent process preceded Kimi K3’s publication, a gap that takes on added significance given the model’s global availability and the scale at which it can now be deployed by any actor with the hardware to run it.