DOC. APF-009·DEPLOY ANY OPEN MODEL·LIVE

Paste a model.
Get an endpoint.

// vs a USD-billing gateway, all-in

−19.6%

Same weights. Same throughput.

Buying the same model in dollars, you pay a 5–7% forex markup and 18% GST on the conversion before the model has run. We hold the provider accounts and bill you rupee-native, so neither exists on your bill.

₹95.5 interbank × 1.06 forex × 1.18 GST vs ₹96.0 flat

▸ See the full comparison

Any supported open model on Hugging Face becomes a private, OpenAI-compatible endpoint in minutes — running on a GPU we rent for you at the cheapest available rate. You pay by the hour, not by the token, and the instance stops itself when nobody is using it.

// Cheaper than a dedicated endpoint. Simpler than renting a GPU yourself.

▸ Deploy your first model See the GPU picker Start free · ₹100

§ 01 — HOW IT WORKS

// Four steps, and you can stop at any of them

From a repo URL to a working endpoint.

Nothing is provisioned and nothing is billed until you have seen the rate and agreed to it.

[01] PASTE

Paste the repo

We check size, licence and architecture before a rupee is spent. If the serving engine cannot load it, you are told then — not after a charge.

[02] SIZE

We pick the card

Model size and serving precision decide the GPU tier, and we find the cheapest card that fits across 30+ clouds.

[03] QUOTE

See the quote, then confirm

A rupee-per-hour rate and an honest first-start estimate — 10 to 45 minutes, because the weights have to download and load. Nothing runs until you say yes.

[04] CALL

Call it like any OpenAI endpoint

Your normal apinfy key, through our gateway, on the OpenAI-compatible path your code already speaks. Idle 30 minutes and it stops itself.

§ 02 — SIZING

// What your model will land on

Size and precision pick the card.

Parameter count times bytes-per-weight, plus room for the KV cache, is what decides which GPU can serve your model. We do that arithmetic for you and then shop it across 30+ clouds.

Model size	Serving precision	GPU tier	Hourly rate
Up to ~4B	bf16 / fp16	Entry (16–24 GB)	Cheapest tier in the catalogue — quoted live
~7–9B	bf16 / fp16	Entry–mid (24 GB)	Cheapest tier in the catalogue — quoted live
~13–15B	bf16, or 8-bit to fit smaller	Mid (40–48 GB)	Mid tier — quoted live
~30–35B	8-bit / 4-bit quantised	Mid–high (48–80 GB)	Mid–high tier — quoted live
~70B	4-bit quantised	High (80 GB)	Highest single-card tier — quoted live

The hourly rate is a live quote, not a rate card. We rent across 30+ clouds and the cheapest card that fits your model changes by the hour, so the rupee figure is shown on the quote screen before you confirm rather than printed here where it would be stale by the time you read it. Tiers are indicative and shift with quantisation and context length.

§ 03 — WHERE IT RUNS

// You choose, and the receipt says so

India, or the cheapest card anywhere.

Most capacity is international, and that is where the cheapest card usually is. If your data has to stay in India, pick an India region at launch — it is priced separately, and the region appears on the quote before the instance starts, not after.

◱ INDIA REGION

Keep it in India

India-resident region available

Choose an India region when you launch. Priced separately from international capacity and not always the cheapest option — the trade is explicit, and the region is stamped on the receipt so you can prove where your data was processed.

◲ INTERNATIONAL

Cheapest card that fits

Region named on your receipt

International spot capacity, which is where the cheapest cards live. We do not claim India residency for it and we do not hide it: the region is on the quote before you start.

§ 04 — LIMITS

// The part most pages leave out

What we will not promise.

HONEST LIMITS · READ BEFORE YOU DEPLOYno asterisks

“Any model” is marketing. Serving engines do not load every architecture. We support 500+ and tell you up front when yours is not one of them — before you are charged.

Gated models need your own token. Llama, Gemma and similar require your Hugging Face token, because the licence is between you and the publisher, not between you and us.

First start takes 10–45 minutes. The weights have to download and load. Restarts are faster. We quote the estimate before you confirm rather than after you have started waiting.

Single-GPU deployments today. Models that need two or more cards are coming. If yours needs them, the check in step one says so instead of failing halfway through a boot.

§ 05 — INTEGRATION

// Drop-in

Your endpoint. Your existing code.

▸ deployed.pyPYTHON 3 // OPENAI-COMPATIBLE

# the same client, the same key — only the model name changes
client = OpenAI(base_url="https://apinfy.com/v1", api_key="kk_live_...")

# whatever you deployed, addressed by its repo id
r = client.chat.completions.create(
    model="deployed/your-org/your-model",
    messages=[...],
)

print(r.receipt.model, r.receipt.region)
# → your-org/your-model  ✓ signed

Paste a model. Get an endpoint.

Billed by the GPU-hour in rupees. Idle instances stop themselves.

▸ Deploy your first model Rent a raw GPU instead