Kimi K3 Launches with Open Weights, Becoming the Largest Open-Source Model Ever Released

Moonshot AI released the weights of Kimi K3 on Hugging Face at 00:00 UTC this Sunday, under a modified MIT license. With a total of 2.8 trillion parameters, it is the largest open-weight model ever disclosed. Only 16 out of the 896 experts in the Mixture-of-Experts activate per token, which reduces the inference cost to that of a dense model of approximately 50 billion parameters, according to the technical documentation published by Moonshot itself.

The native download in MXFP4 is approximately 594 GB. Requantizations in the community's GGUF Q4 are expected to reduce this to between 300 GB and 400 GB in the upcoming days. Running the model requires between 4 and 8 H100 80 GB GPUs, which places self-hosting beyond the reach of nearly all companies without a dedicated cluster, but within reach for banks, big tech companies, and regional cloud providers.

The Kimi K3 has been operational since July 16 via its own API and in the consumer app kimi.com, when Moonshot announced it at the World AI Conference in Shanghai. The release of the weights now facilitates inspection, fine-tuning, and private execution.

The Timing of the Launch is No Accident

Two days prior, on July 25, Jensen Huang, CEO of Nvidia, published a manifesto urging the U.S. to lead in open weights. Sam Altman responded on X: 'I want the U.S. to win in AI in both open source and proprietary models, and I'm glad to see it.' Sundar Pichai wrote that Google 'supports this on behalf of Google. We have always benefited from open source, we are major contributors and have consistently made open weights available for Gemma.' Elon Musk expressed support for the cause. Anthropic and Amazon did not sign.

The sequence matters: 24 hours after the American coalition called for leadership in open weights, China released the largest open-weight model in history. It is a snapshot of the asymmetric race of 2026, where Beijing uses open weights as a competitive weapon while Washington still debates whether openness is a strategic asset or a security risk.

Where the Model Lands

In the United States, Kimi K3 poses a direct threat to the revenue model of hyperscalers. Moonshot’s API pricing is a fraction of the price of Claude Opus 5 ($5 input, $25 output per million tokens) and GPT-5.6. A company that currently pays millions per year for Claude or OpenAI tokens now has the option to run a comparable model in coding and reasoning within its own VPC.

In Europe, Kimi K3 arrives just six days before the enforcement of Article 50 of the EU AI Act, set for August 2, 2026. Fines for non-compliance with the obligations for transparency of synthetic content can reach up to 15 million euros or 3% of global revenue. A model with available weights facilitates compliance because the auditor can inspect behavior, training data, and watermarking mechanisms. A model only available via API cannot.

In Japan, where MUFG, Mizuho, and Sumitomo have already negotiated access to OpenAI’s new model according to Nikkei, Kimi K3 adds an alternative that resolves data sovereignty issues without relying on contracts with Microsoft or Amazon. This is the same calculation MUFG made on July 16 when it invested in Noetra to develop domestic foundation models.

In Brazil, three fronts open up. Banks such as Itaú Unibanco, with over 500 internal AI use cases in production, gain the chance to run a frontier model within their own data centers without violating LGPD or Circular 3,909 from the BCB regarding cloud computing. Consultancies like CI&T and Falconi now have a competitive argument against the OpenAI/Anthropic oligopoly. And shared services hubs serving global clients can offer AI services from infrastructure installed on Brazilian soil.

There are legitimate counterarguments. The first is the hallucination rate: Moonshot has not published hallucination benchmarks for K3, although it has released competitive scores in coding and agentic use. The second is the risk surface: running a cutting-edge Chinese model in production entails audit from weight to weight to detect unwanted behaviors, and this engineering costs more than an average company realizes. The third concerns geopolitical issues: there is no guarantee that the modified MIT license will not be revised in a future iteration.

Despite these caveats, Kimi K3 is the landmark marking that frontier AI is no longer the exclusive property of a handful of American labs. In the coming weeks, we will see who can run the model in production before the community quantizes practical versions for less extravagant hardware.