Kimi K3 open weights

What's official, what's unknown

Anyone quoting a minimum-VRAM figure, a serving command, or a license clause for K3 today is guessing. The honest version of this page is a short one until July 27 — what we can do honestly is arithmetic.

The hardware math

2.8 trillion parameters have a fixed storage cost that no serving trick removes. This is multiplication, not measurement:

Two things follow from the arithmetic. First, the MoE sparsity doesn't shrink the footprint. Activating 16 of 896 experts cuts the compute per token, not the memory: every expert has to be resident for any of them to be routable. Second, no single machine holds this. Even at 4-bit, 1.4 TB of weights is roughly triple the largest unified-memory workstation and an order of magnitude past any single GPU — this is a multi-node deployment on day one, before you budget the KV cache for a 1,048,576-token window.

For scale: K3's 2.8T is 2.8× the footprint of the 1T-parameter K2 family at the same precision — and the K2 releases were already multi-GPU deployments. Real serving requirements (tensor-parallel layout, supported engines, quantized builds) arrive with the model card; we'll replace this arithmetic with the official numbers the day they exist.

Ready-for-the-drop checklist

• Watch the moonshotai Hugging Face org. Moonshot hasn't said where the weights will land, but every prior Kimi open release has been published there.
• Read the license before you build. It is not announced. Don't assume it matches earlier Kimi releases until the text is public.
• Hold your serving-stack decisions. KDA is a new attention design — whether your inference engine supports it on day one is exactly the kind of thing the model card answers and speculation doesn't.
• Price the deployment honestly. If the 4-bit row above is out of budget, self-hosting K3 isn't your path — and that's a fine outcome, because the hosted model is one base_url away.
• Baseline the model now. The API is live: you can validate quality, latency, and cost on your actual workload before committing to hardware. Our measurements are a starting point.

Self-hosting vs. the API, in numbers

The API's entry ticket is $0. The self-hosted entry ticket is the 1.4 TB row above. Between those extremes, the numbers we've measured give you the API side of the ledger with unusual precision:

• A 497,718-token prompt costs $1.49 in one call at the official $3.00/1M rate — measured, and a full-window call works out to about $3.15.
• Prompt caching cuts repeated input to $0.30/1M automatically — the 10× discount that makes "same big context, many questions" workloads viable without owning a single GPU.
• Your workload's monthly bill at official rates: the calculator computes it in your browser. If that number is a small fraction of a multi-node GPU cluster's monthly cost, the decision has made itself — self-hosting K3 is for compliance, data-residency, and research reasons, not for saving money at typical volumes.

What you can run today

The hosted model is live and speaks the OpenAI format. This is the exact code we verified against production the day K3 launched:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_EVOLINK_API_KEY",
    base_url="https://direct.evolink.ai/v1",   # <-- the one line you change
)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Explain prompt caching in two sentences."}],
)

print(response.choices[0].message.content)

Auth, streaming, and what a real response body looks like: the API guide. Leave max_completion_tokens at its default — here's why.

Open-weights FAQ

When will the Kimi K3 weights be released?

Moonshot AI's official documentation commits to "by July 27, 2026" (read 2026-07-18). As of 2026-07-18 the weights are not yet published. This page is updated the day that changes.

Is Kimi K3 open source?

Committed but not yet delivered: the weights are officially due by July 27, 2026, and no license has been announced. Until both the weights and the license text are public, K3 is available only as a hosted model.

What hardware does it take to run Kimi K3 locally?

Unknown until the model card ships. The arithmetic floor: 2.8T parameters is 1.4 TB at 4-bit — at least eight 192 GB accelerators just to hold the weights, before KV cache for the 1M-token window. It will not run on a single GPU or a single workstation at any published precision.

Where will the weights be downloadable?

Not officially stated. Every prior Kimi open release landed on the moonshotai Hugging Face organization, which makes it the place to watch.

Can I use Kimi K3 before the weights drop?

Yes — the API has been live since launch. Model ID kimi-k3, OpenAI-compatible, $3.00/1M input and $15.00/1M output at the official rates. First call in five minutes.