The biggest open-weight model ever shipped today, and you probably can''t run it

Moonshot AI released Kimi K3 this morning. Full open weights, 2.8 trillion parameters, roughly 1.4 terabytes on disk even after MXFP4 quantization. It landed in the same week DeepSeek V4 went stable. The last stretch of July has produced the heaviest concentration of open-weight releases the industry has ever seen.

Every headline says "open." For most teams the honest translation is "look, but don''t touch."

"Open weight" stopped meaning "you can host it"

Downloading the weights is the easy part. Serving them is where reality bites. A single H100 holds 80GB of VRAM. A 1.4TB model needs something like 18 of those cards just to hold the weights in memory, before you add KV cache, batching headroom, or any concurrency. That is a multi-node cluster with fast interconnect, an inference stack tuned for a mixture-of-experts model, and someone on call when a node drops mid-request.

This is not a weekend Docker project. It is a small platform team with a real budget.

The self-hosting math rarely works anyway

Even for models you can fit on a box, the numbers usually point the other way. Break-even for self-hosting a mid-size open model lands somewhere between 10 and 30 million tokens a day. For a frontier model this size, that break-even runs into billions of tokens a month, which most teams never touch.

Then there is the part nobody budgets for. A senior inference engineer runs $250K to $360K loaded in the US. If your closed-API bill is $8K a month, you cannot justify a $25K-a-month hire to bring it in-house. The engineer costs more than the thing you were trying to save on, by a factor of four or five.

Meanwhile budget open-weight APIs (Together, DeepInfra, Fireworks, DeepSeek''s own endpoint) are sitting around $0.14 to $0.50 per million tokens. That is hard to beat with your own GPUs unless you are running at a scale that comes with its own headaches.

So what do open weights actually buy you

Leverage, not independence. Because the weights are public, many providers race to serve the same model. That competition pushes prices down and gives you an exit if one provider disappears or a lab decides to gate its hosted version. Your safety net is portability, not a rack in your own datacenter.

Here is how we think about it at Devdot when we build with these models:

Consume frontier open models through an inference provider, and keep your integration layer provider-agnostic so you can switch in an afternoon.
Reserve actual self-hosting for the small models, where privacy, latency, or offline operation matter more than raw capability.
Route by task. Cheap or self-hosted models for routine traffic, a frontier API for the hard requests. That hybrid split routinely cuts spend by 30 to 50 percent.

Kimi K3 being open is genuinely good news. It keeps the hosted market honest and hands you real negotiating power. Just don''t read "open" as "run it yourself." For a 1.4TB model, that instinct will cost you a cluster and an engineer you didn''t need.

We''re here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you''re looking to build something, get in contact with us today!