Kimi K3's Open Weights Are the Largest Ever Released

What does "open weights" actually buy you when the download runs to 1.4 terabytes? That is the question worth sitting with after Moonshot AI uploaded the weights for Kimi K3 to Hugging Face on July 27, making it the largest AI model whose weights have ever been published.

At 2.8 trillion parameters, K3 arrived with more than just the checkpoint. Moonshot also released supporting infrastructure: high-performance attention kernels, a communication library for mixture-of-experts models, and tooling for running AI agents at scale.

How much of 2.8 trillion is actually running

Not all of it, and the gap matters. K3 uses a mixture-of-experts architecture with 896 expert sub-models, of which only 16 activate per token. That works out to roughly 104 billion active parameters for any given token. On paper it is a 2.8 trillion parameter model; in terms of what it costs to run, it behaves closer to a 104 billion parameter one.

The context window is 1 million tokens. Moonshot says a new attention mechanism it calls Kimi Delta Attention makes long-context inference up to six times cheaper than previous approaches. The weights ship in MXFP4 format, meaning the quantization was accounted for during training rather than bolted on afterward.

Where it lands on benchmarks

Independent measurements put K3 near the top without placing it at the summit. On the Artificial Analysis index it ranks 4th out of 189 models. Coding is its strongest showing, with 88.3 on Terminal-Bench 2.1. It falls noticeably behind frontier models on cybersecurity tasks and mathematics.

The release also arrived in the middle of a political argument. Washington claimed earlier this month that the model had been distilled from Anthropic's Fable, a charge outside experts disputed. That dispute is unresolved, and it is worth separating from the technical question of what the weights can do.

The license fine print

The weights carry a "Modified MIT" license. Commercial use, modification and redistribution are permitted with attribution, but two thresholds apply. Companies that serve the model as a service and earn over $20 million a year need a separate agreement with Moonshot. Products above 100 million monthly users or $20 million in monthly revenue must display "Kimi K3" in their interface.

Neither threshold binds the vast majority of businesses. The point is that open weights and open source software licensing are not the same category, and assuming otherwise is how legal surprises happen two years into a product.

Running it yourself is the expensive part

The license is not the real barrier. Serving K3 requires more than 1.4 terabytes of GPU memory, and Moonshot recommends upwards of 64 accelerators for a production deployment. It does not fit on one server, or even on a single 8-GPU node. You need a multi-node cluster, which puts self-hosting out of reach for all but a small number of organizations.

So for most companies this is not a self-hosting story at all. It is a pricing story. Moonshot's own API runs $3 per million input tokens and $15 per million output, dropping to $0.30 on a cache hit, well under the closed frontier models. And because the weights are public, cloud providers can now serve K3 on their own infrastructure, which widens the options for organizations with data residency requirements.

Two questions before you switch anything

The reflex when a model like this lands is to ask whether to migrate. That is the wrong first question. Two better ones:

• Is this model strong at your work? K3's advantage shows up in coding and long documents. If your workload is mathematical reasoning or security analysis, the benchmark story reverses.
• Through which channel will you consume it? Hosted API, cloud provider, or your own cluster are three very different cost and compliance profiles for the same weights.

There is also a third option that gets overlooked. In a test by Fireworks AI, a router splitting work between Kimi K3 and a frontier model achieved 93% accuracy. Committing to one model is rarely necessary, and sending each task to the model that suits it usually costs less and performs better than either choice alone.

Sources: The Decoder, VentureBeat, Hugging Face