Self-hosted Kimi K3 · Founding reservation
Kimi K3 at half the public API price—served directly from GPU infrastructure we operate, without a router or third-party inference API.
Community-funded launch
Launches immediately at $10,000. Full refunds if not funded by August 30.
Why I’m building this
I’m Dhrumil—one heavy API user, not a funded startup. Kimi K3 is open-weight, and its license permits people to run and deploy the model. That means I can operate the weights on rented GPUs instead of buying Moonshot’s hosted API, so their API price does not have to become your price.
Free signups are easy to collect, but they do not tell me whether enough people will actually use the service. Paid reservations measure real demand. If we reach $10,000, I will use the committed funds to reserve the inference capacity needed to launch.
Your payment remains tied to your Kimi K3 reservation. It cannot be spent on existing Vynaris services. If the target fills, it converts 1:1 into account credit when inference launches. If we do not proceed, every reservation is refunded to its original payment method.
Launch pricing
Every price is per one million tokens. No blended router price and no separate platform markup.
Launch traffic will use Kimi K3’s native MXFP4 weights and MXFP8 activations, with no additional down-quantization. If people ask for a lower-precision version at a lower price, we will offer it later as a separate and clearly labelled tier.
Public-price comparison checked July 31, 2026 against the Kimi API list price. “Cheapest” will be reverified before reservations open. Cached input, taxes, and future promotional prices are excluded.
Private by architecture
Specific-model requests bypass Vynaris Auto. They go straight to the Kimi K3 deployment running on GPU capacity we operate.
Your request is not forwarded to an external model provider.
Request content is processed for inference, then discarded.
Token counts, costs, and operational records—not your content.
How the reservation works
Make one reservation or add more later. There is no per-account limit.
You receive a payment receipt, but the amount is not spendable before launch.
If the target fills, we launch immediately and it becomes Vynaris credit. If not funded by August 30, it returns to your original payment method.
One familiar API
Keep your OpenAI-compatible client. Set the model explicitly and Vynaris skips automatic routing.
from openai import OpenAI
client = OpenAI(
api_key="vyn_...",
base_url="https://api.vynaris.com/v1",
)
response = client.chat.completions.create(
model="vynaris/kimi-k3",
messages=[{"role": "user", "content": "..."}],
)Before you reserve
A paid launch reservation. Your money is held as locked Kimi K3 credit, not added to your spendable Vynaris balance yet.
I am an individual, not a funded startup. A paid reservation shows real demand and gives me the money needed to reserve dedicated GPU capacity without raising outside funding.
If the target is not reached by August 30, 2026, your full reservation is refunded to the original payment method. If it fills earlier, we begin the launch immediately.
Yes. There is no per-account reservation limit. Every successful reservation is added to your total locked credit.
After Kimi K3 launches, the reservation becomes normal Vynaris credit. You can use it on Kimi K3 or any other available model.
It is the introductory launch rate, not a lifetime guarantee. Any future pricing change will be communicated before it takes effect.
No. Kimi K3 API prompts and completions will not be retained. We keep only the account, token, cost, and operational records needed to bill and run the service.
No. The launch service will run Kimi K3 at its native released precision—MXFP4 weights with MXFP8 activations—with no additional down-quantization. If users later want a cheaper lower-precision option, it will be offered as a separate, clearly labelled tier.
Founding access
$50 minimum. Immediate launch at $10,000—or a full refund after August 30.