Pakistan’s Sovereign AI Gateway

Access leading AI models in Pakistan, with sovereignty and control

One LLM gateway in front of every open-source model we serve: fast, cost-effective inference, billed in rupees.

Powered by optimised open-source AI

One API. Every leading open-source AI model.

Every route P/9 serves today, priced per million tokens in rupees. Every card is the same catalogue the dashboard bills against.

Global

gpt-oss 120B

OpenAI · 117B total, 5.1B active

Chat completions · 131,072 tokens context

Input / 1M

Rs 30.08

Output / 1M

Rs 180.47

Tap to open the model card

Global

Kimi K3

Moonshot AI · 2.8T total, 104B active

Chat completions · 1,048,576 tokens context

Input / 1M

Rs 902.34

Output / 1M

Rs 4,601.93

Tap to open the model card

Global

GLM-5.2

Z.ai

Chat completions

Input / 1M

Rs 421.09

Output / 1M

Rs 1,365.54

Tap to open the model card

Global

DeepSeek V4.1 Flash

DeepSeek AI

Chat completions

Input / 1M

Rs 90.23

Output / 1M

Rs 363.04

Tap to open the model card

PK NPU

DeepSeek-OCR 2

DeepSeek

Chat completions

Input / 1M

Rs 9.02

Output / 1M

Rs 9.02

Tap to open the model card

View all models
1. NPU-native inference in Pakistan

Huawei NPUs are going to get better and be everywhere.

There is a learning-curve challenge to using Huawei's Ascend NPUs.

But Pakistan, along with other countries that face Nvidia import challenges, will mostly have NPUs in the country.

We are here to bridge that gap. We give you highly optimised open-source model endpoints, with a sovereign gateway that works on both Huawei NPUs and Nvidia GPUs.

Dedicated support, and accountability for uptime, on Huawei NPUs.

Huawei Ascend 910B
64GB NPUs, racked in country rather than rented by the hour in a foreign region.
One gateway, two kinds of silicon
The same route id serves from Ascend NPUs or Nvidia GPUs. Your code does not know which, and does not need to.
Accountable uptime
The people who wrote the Ascend serving stack are the people on call for it. Dedicated support, not a ticket queue.

Already built on CUDA

CUDA
CANN

CUDA to CANN translation.

If your model is CUDA-native and already prepared for deployment on GPUs, that is not a blocker. We translate the model code to run on CANN, Huawei’s Ascend compute stack, and serve your inference on NPUs with the performance numbers to show for it.

“P/9 Gateway”Sovereign gateway features

Give your whole team access to every model, then see, control and cap every request.

01

Access: give your whole team every model.

Put every route P/9 serves behind one OpenAI-compatible API and one organisation login. Keys are minted per project and scoped to the models you choose, and the alias you call stays stable when the weights behind it move, so platform teams open access without becoming the bottleneck.

  • One OpenAI-compatible API to every route. Change the model string, not your stack.
  • Keys per project, scoped by allowed models, with an expiry and a rupee credit limit.
  • PK NPU-native and global pass-through routes through the same gateway and the same keys.
  • New open models arrive as routes, not as a code change on your side.
02

Visibility and control: see every request, and cap it before it runs.

Usage and spend are attributed to every key, project and model, so chargeback is a filter rather than a spreadsheet. Give a key a hard credit limit and it stops at the cap with a 402 while the rest of your balance stays intact. Telemetry, the audit log and the ledger are kept in Pakistan.

  • Requests, tokens, time to first token and throughput per key, per project and per model.
  • Hard credit limits per key: at the cap that key stops, and nothing else does.
  • Expiry and instant revocation, honoured by the gateway on the very next request.
  • Model access control per key, with a 403 for anything outside scope.
  • An append-only audit log of every administrative change, exportable as CSV.
03

Cost: priced in rupees, end to end.

Rates are quoted per million tokens in PKR, debited from a prepaid balance and settled in the same currency. No foreign card, no exchange-rate surprise at month end, and the exact cost of every request reported back by the gateway.

  • Per-million-token prices in rupees, visible before you call.
  • A prepaid balance: budget exhaustion fails closed, never on credit.
  • Per-key spend for chargeback across teams and products.
  • The cost of each request returned in the response's debug metrics.
04

Performance: two milliseconds of overhead. Read the benchmark.

The gateway is a Go service on the request path: it authenticates the key, checks scope and budget, meters the tokens and relays. Its overhead is measured with an open, reproducible harness against a deterministic mock upstream, beside four other gateways on the same host, so the number is the gateway's alone and the comparison is a fair one.

  • About 2 ms added at p50 and under 7 ms at p99, key, scope and budget checks included.
  • 4,520 requests a second sustained from two vCPUs, ahead of every gateway measured beside it.
  • A 3.4 MB image and under 50 MB of memory, so it runs beside the model rather than on a fleet of its own.
  • Run it yourself: the harness is open source, one command on a fresh box. See the harness (opens in a new tab)
2. Intelligent model routing

The P/9 Router predicts and sends each request to the best-fit open-source model, cutting cost while holding quality.

Incoming request

Simple chat

P/9 Router
  • p9-fastQwen3-8B
  • p9-coderQwen3-Coder
  • p9-reasongpt-oss 20B
  • p9-largeDeepSeek V3.2
  • p9-embedbge-large

Per-request routing

A P/9 agent reads each prompt and task, then routes to the best-suited model in the fleet.

Quality preserved

Top-tier answers by sending hard requests to stronger models, easy ones to lighter, cheaper models.

Cost driven down

Most traffic is served by lightweight open-source models, so you only pay for heavyweight compute when it matters.

3. One unified API

One OpenAI-compatible API, connected to optimised open-source models on sovereign infrastructure.

api.paranine.com/v1
# drop-in with any OpenAI-compatible SDK
from openai import OpenAI
base_url= "https://api.paranine.com/v1"
api_key= "sk_p9_•••••••"
model= "p9-coder"
# served on sovereign compute · billed locally in PKR
Drop-in OpenAI-compatibleSovereign infrastructureBilled locally
Pricing

Only pay for what you need.

Just pay for the API calls you make, with a transparent platform fee. New organisations start with Rs 1,000 of free credits. Every price is in PKR: quoted, billed and settled in the same currency, with no exchange-rate surprise at the end of the month.

LLM Gateway

Pay-as-you-Go

Top up your account and pay per token usage as per models per-million-token rate.

Browse models

Included

  • Access to every P/9-hosted model with one unified, OpenAI-compatible API
  • Compare models on published pedigree, latency and price
  • Multiple API keys for different projects, each with its own scope and cap
  • Cost and performance monitoring, per key and per model
  • Chat support (48h, working days)
  • Unlimited seats

Enterprise Gateway Platform

Custom price

Quoted per organisation, for dedicated capacity and enterprise terms.

Talk to sales

All the previous features +

  • High rate limits and bulk discounts
  • Private deployments and specific compliance needs
  • Custom latency and cost optimisation
  • Dedicated support, SLA and professional services
  • Custom billing and invoicing structures
  • Pay as you go
  • No hidden costs
  • No API call limit

Designed for production-grade workloads

1. Our inference is state-of-the-art for lightweight open-source models, with higher accuracy and cost efficiency, powered by in-house research.

2. Integrations are stack-agnostic through one secure, OpenAI-compatible API. Recommendations execute in your gateway and harness of choice.

3. Cost tracking, spike alerts, error and latency monitoring, eval tooling, local data residency and 24-hour user support.

Sovereign infrastructureLocal data residencyOpenAI-compatible24-hour support

Common questions

Running AI in Pakistan, answered.

Is there an AI API that bills in Pakistani rupees?
Yes. paranine is billed entirely in PKR from prepaid credits, at a rate published per million tokens for each model. There is no dollar conversion, no card in a foreign currency and no exchange-rate exposure between the day you top up and the day you spend it.
Where is my data processed when I call the P/9 API?
Requests to the PK NPU-native pool are served on P/9's own Huawei Ascend hardware inside Pakistan. The gateway stores no prompts and no completions on any route: a request log keeps the model, key, status, timings, tokens and cost, and nothing of what was said.
Is paranine OpenAI-compatible?
Yes. The gateway serves the OpenAI chat completions and embeddings shapes, so any OpenAI SDK works by changing two values: the base URL and the API key. Anthropic's native Messages API is served on the same host for clients that speak it.
How is this different from OpenRouter, Together AI or DeepInfra?
Those route to capacity wherever it is cheapest, which is a good answer if location does not matter to you. P/9 serves open models from inside Pakistan on hardware it operates, bills in rupees, and tells you which pool served each request, which is the answer when data residency or a local invoice is a requirement rather than a preference.
Which models can I call, and what do they cost?
The catalogue lists every route P/9 serves with its price per million input and output tokens in rupees. It covers open-weight models on global pass-through routes and on the PK NPU-native pool, and the same page marks which pool serves which route.
Is there an LLM gateway in Pakistan?
Yes. paranine operates one: a single OpenAI-compatible API in front of many open-weight models, with routes served on Huawei Ascend hardware P/9 runs inside Pakistan alongside global pass-through routes, billed in rupees from prepaid credits. Keys carry a rate limit, a rupee spend cap, an expiry and a model scope, all enforced by the gateway.
Do I need my own GPUs to run AI in Pakistan?
No. Nvidia import constraints mean most accelerator capacity reaching Pakistan is Huawei Ascend, which is a different toolchain from CUDA. A gateway is how a team gets that capacity through an OpenAI-compatible endpoint without any application learning CANN.

More on how this works: what a sovereign AI gateway is, the model catalogue and its rupee prices, and what the PK Endpoint commits to.

Cost-effective inference for every leading open model.