Compatibility
Bring your HTTP client
Every open-weight models answers on one endpoint - POST https://https://api.deepseek.com/v1 Drop in the key and the model id, there is no extra library to install.
open-weight token packs
Prepaid packs for every open-weight family - Astra, Sol, Mini and Luna. Fixed token count, per-million price, no meter surprises.
Every open-weight pack is a virtual product. There is no shipping, no stock limit and no waiting. Buy 10+ units of a pack for 5% off, or 50+ units for 10% off. Prices are shown in your selected currency at checkout.
Open the full model directory →Model directory
open-weight, open-weight Sol, open-weight and open-weight, plus batch-rate and prompt-caching packs, all sold as token packs from one account with one checkout. Compare them here, then add the pack you need to the cart.
Compatibility
Every open-weight models answers on one endpoint - POST https://https://api.deepseek.com/v1 Drop in the key and the model id, there is no extra library to install.
Billing
You pay for a fixed quota up front, so there is no monthly invoice, no auto-recharge and no overage.
Delivery
There is no shipping address and no delivery wait. The key appears in your account and inbox once the order completes.
Traceability
Every key is tied to an order line, a model and a quota, which makes spend easy to audit.
Fit check
Tired of topping up balances and chasing invoices across projects? Buy the exact quota you need as a pack instead.
| Buying credit directly | With TokenHub packs | |
|---|---|---|
| Accounts & keys | Create an organisation, set up billing and manage project keys yourself. | Checkout once. A key for the model arrives with the order. |
| Budget control | Watch usage dashboards and set spending limits after the fact. | Quota is fixed by the pack, so a key cannot overspend. |
| Payment | A card on file, billed monthly or topped up by auto-recharge. | Pay per order with the methods offered at checkout. |
| Delivery | Credits applied to the whole organisation. | One key per order line, with its own model and quota. |
Order + key trace
Your checkout stays the same. The pack you pick, the number of units and the order status decide how much quota the key carries and whether it is active.
| Step | Single pack | Stack units | Refund |
|---|---|---|---|
| Order | the model · 1 unit | the model · 12 units | the model · 1 unit |
| Checkout | No shipping step | 10+ units −5% | Complete |
| Status | Complete | Complete | Refunded |
| Key | sk-••••••3f9a | One key per line | Suspended |
| Quota | 2M tokens | 24M tokens | Back to Complete |
Pack prices already include the volume breaks for 10+ and 50+ units. Flagship packs currently carry up to 15% off list price. Check the model page for the live price in your currency.
First call
Registration is free and does not include model credit. Quota comes from the packs you buy.
Register with an email. Keys are stored under your order history.
Registration is freePick the model and the number of units, then pay at checkout. There is no shipping step.
Volume breaks at 10+ and 50+ unitsOnce the order completes, copy the key from My Account › Order History and send your first request.
Key issued automaticallyimport os, requests
resp = requests.post(
"https://https://api.deepseek.com/v1",
headers={
"Authorization": "Bearer " + os.environ["the platform_API_KEY"],
"Content-Type": "application/json",
},
json={
"model": "YOUR_MODEL_ID",
"input": "Explain the CAP theorem in two sentences.",
},
)
print(resp.json()["output"][0]["content"][0]["text"])curl https://https://api.deepseek.com/v1 -H "Authorization: Bearer $the platform_API_KEY" -H "Content-Type: application/json" -d '{
"model": "YOUR_MODEL_ID",
"input": "Explain the CAP theorem in two sentences."
}'const res = await fetch("https://https://api.deepseek.com/v1", {
method: "POST",
headers: {
"Content-Type": "application/json",
Authorization: `Bearer ${process.env.the platform_API_KEY}`,
},
body: JSON.stringify({
model: "YOUR_MODEL_ID",
input: "Explain the CAP theorem in two sentences.",
}),
});
const data = await res.json();
console.log(data.output[0].content[0].text);Quota is attached to the key, not to your account - a request never draws from a shared balance.
Pricing, without the guesswork
Each model is sold as a fixed token pack, priced per million tokens at Hugging Face's own rate, and you can stack as many units as you need.
Batch API packs are billed at half the standard rate, and prompt-caching packs bill cached input at 10% of the input price with 1.25x writes.
5% off from 10 units and 10% off from 50 units. The discount is applied automatically.
Prices convert to the currency picked in the header. You pay the amount shown at checkout.
Use any payment method the checkout page offers for your order.
Data handling
The store handles your order and the key it issues. The prompts and responses you send with that key never pass through the shop.
Order id, items and status
Issued access key and its model id
Remaining balance attached to the key
Never received by the storefront
Virtual delivery
Token packs are virtual products. There is no warehouse and no courier involved. The store issues, suspends and restores access keys from the order status.
Token packs never ask for a delivery address or a shipping method.
A key is generated for each virtual order line once payment completes.
If an order leaves a completed status, its keys are suspended.
Keys, model ids and quota live under My Account › Order History.
Company
Kimi TokenHub sells prepaid GPU-hour packs for open-weight models published on the Hugging Face Hub.
Every product is one open-weight model plus a block of GPU hours: StarCoder2, Qwen2.5-Coder, MetaMath, Phi-2, BART, mBART, RWKV-5, NLLB-200, wav2vec2 XLSR, E5 Large, Jina v3 and Mamba.
T4 16GB at $0.50/hour, L4 24GB at $0.80/hour, A100 80GB at $2.50/hour. Fifty-hour packs carry a 10% volume discount.
The endpoint key appears in My Account / Order History once the order completes.
Unused packs can be refunded within 14 days while no request has been made.
Ready to start?
open-weight token packs from Hugging Face, delivered as virtual products.