Author(s): Ashish Nishad

Originally published on Towards AI.

Kimi K3 has 2.8 trillion parameters, taught itself to design a chip to run a smaller version of itself, and paused new signups three days after launch because it couldn’t handle demand. Here’s what it actually is — and what happened when I put it to work.

On July 16, a Beijing-based lab called Moonshot AI released a model with 2.8 trillion parameters — by multiple independent accounts, the largest open-weight AI model ever built. Three days later, Moonshot paused new signups because demand had pushed the system close to its processing limits. It arrived days before the 2026 Artificial Intelligence World Conference in Shanghai, and it’s called Kimi K3.

After the intro, the author explains what’s new about Kimi K3 beyond its headline size: it’s a Mixture-of-Experts model with selective activation, specific architectural changes aimed at efficiency gains, quantization-aware training for better hardware compatibility, a very large context window, and image input—plus a “designed-chip” proof of concept where K3 autonomously produced a tiny runnable chip design. The article then discusses the benchmarks with important caveats about inconsistent harnesses and occasional vendor-run tests, while noting K3’s strong performance across several coding evaluations and generally near-frontier capability at the cost of higher token usage and slower generation. The author compares these results with their own hands-on code-generation experience, describing alignment between real tasks and the published metrics, and outlines practical limitations (open weights weren’t immediately verifiable at launch, it’s primarily an API/enterprise model given the scale, and pricing is premium but not extreme). Finally, the author argues why this matters more broadly for Chinese AI development—framing K3 as evidence that architectural innovation plus large-scale training can overcome compute constraints and reinforcing the developer value of the price-to-capability trend.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI