LLMnow.ai — GPU-Powered LLM Inference for Turkey & MENA

One line. That is the migration.

LLMnow speaks the OpenAI API. Point your existing SDK at our base URL and keep every other line of code.

Chat completions, streaming, function calling and embeddings follow the same schema you already use.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.llmnow.ai/v1",
    api_key="llm-...",
)

stream = client.chat.completions.create(
    model="llama-3.3-70b-instruct",
    messages=[{"role": "user", "content": "Merhaba!"}],
    stream=True,
)

for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.llmnow.ai/v1",
  apiKey: process.env.LLMNOW_API_KEY,
});

const stream = await client.chat.completions.create({
  model: "llama-3.3-70b-instruct",
  messages: [{ role: "user", content: "Merhaba!" }],
  stream: true,
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

curl https://api.llmnow.ai/v1/chat/completions \
  -H "Authorization: Bearer $LLMNOW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama-3.3-70b-instruct",
    "messages": [{"role": "user", "content": "Merhaba!"}],
    "stream": true
  }'

The problem

Three problems. One region. Zero solutions.

Latency you can feel

A round trip from the region to a US inference endpoint adds 220ms or more before a single token is generated. In chat and agent workloads that cost compounds on every turn.

KVKK exposure

Global providers cannot guarantee where your prompts are processed or stored. For regulated data, sending it abroad is a compliance decision you have to defend — not a technical detail.

FX risk on every invoice

USD-denominated API bills turn an engineering cost into a currency position. Budgets set in January stop meaning anything by June.

“We need AI, but we cannot send customer data abroad.”
— CFO, Istanbul fintech

The platform

Regional inference, without the trade-offs.

Everything you expect from a managed inference provider, hosted where your users and your regulator both are.

01

Our data centres in Türkiye

RTX PRO 6000 Blackwell (96 GB) clusters alongside Ampere capacity — more than 1.3 TB of GPU memory, all located in Türkiye. Short, predictable network paths for users across Türkiye, MENA and Europe.

02

Data residency

Prompts and completions are processed in Turkey. A data processing agreement is included on every paid plan, written against KVKK.

03

50+ open models

Llama, Mistral, DeepSeek, Qwen and Turkish fine-tunes. Bring your own weights and we will serve them on dedicated capacity.

04

Drop-in compatible

OpenAI-compatible REST API. The official OpenAI SDKs, LangChain, LlamaIndex and Vercel AI SDK work unmodified.

05

Auto-scaling

Traffic spikes are absorbed by the shared pool. Reserved capacity is available when you need a guaranteed floor.

06

Billing in TRY

Priced and invoiced in Turkish Lira with e-fatura. USD pricing remains available for international customers.

Coverage

Infrastructure in Türkiye. Service across three regions.

Our GPUs run in data centres in Türkiye. From there we serve Türkiye, MENA and Europe — and your data never leaves the country.

Türkiye

Infrastructure here

Blackwell and Ampere clusters, KVKK data residency and invoicing in Turkish Lira. Lowest latency of the three.

MENA

Served

The Gulf and the wider Middle East. USD invoicing, English and Arabic interfaces. A markedly shorter network path than US-hosted providers.

Europe

Served

EUR/USD invoicing for European teams. Suited to organisations that want their data outside the EU but in a jurisdiction they can audit.

Regions on the roadmap

These are planned regions — we do not serve from them today. Each will be listed with its own legal regime when it opens.

Frankfurt EU / GDPR planned
Dubai Gulf, low latency planned
Bulgaria EU, low cost planned
Canada North America / PIPEDA planned

Need a specific region sooner? Say so in the early access form — we sequence by demand.

Coverage describes the regions we serve; all GPUs are located in Türkiye. If you need a local deployment in a specific country, let us discuss it on an enterprise plan.

Models

Open weights, served fast.

The catalogue below is the launch set — models we can serve on the hardware we operate today. Per-model pricing and context limits are published when the API opens.

Llama

Llama 3.3 70B Instruct
Llama 3.1 8B Instruct
Llama Guard 3 8B

Qwen

Qwen2.5 72B Instruct
Qwen2.5 Coder 32B
Qwen2.5 7B Instruct

DeepSeek

DeepSeek-R1 Distill 70B
DeepSeek-R1 Distill 32B

Mistral

Mistral Large
Mixtral 8x7B Instruct
Mistral 7B Instruct

Turkish

Turkish-tuned chat models
Turkish embedding models
Bring your own fine-tune

Utility

Text embeddings
Reranking
Speech-to-text

Model availability at launch depends on licence terms and cluster capacity. Need something specific? Tell us in the waitlist form.

Use cases

Which model for which job

Choosing a model is a cost decision, not just a technical one. Below are the workloads we see most often, the model we recommend for each, and a realistic monthly figure.

Customer support assistant

50,000 conversations/mo ≈ $62

FAQs, order status, returns. The first point of contact before a human takes over.

Model: Llama 3.1 8B · Qwen2.5 14B
Context: 8K

Less than a tenth of what one support agent costs per month.

Document and archive search (RAG)

200,000 queries/mo ≈ $70

Internal documents, procedures, technical manuals. Staff ask in plain language and get an answer with its source.

Model: Embeddings + Mistral Small 24B
Context: 4K–8K

Embedding and reranking run on the same platform; nothing leaves the country.

Call and meeting summarisation

20,000 calls/mo ≈ $18

Summaries, action items and customer intent extracted from a transcript.

Model: Qwen2.5 14B · Llama 3.1 8B
Context: 8K

Speech-to-text can run on the same platform.

Contract and compliance review

500 contracts/mo ≈ $15

Finding risk clauses in long text, comparing versions, producing summaries. For legal and procurement teams.

Model: Llama 3.3 70B · Qwen2.5 72B
Context: 32K

Our most capable models — and the document never leaves the country.

Coding assistant

20 developers ≈ $8/mo

Completion, review, test generation. Without your code or repository leaving your control.

Model: Qwen2.5 Coder 32B
Context: 8K–32K

For projects where source code cannot be sent abroad.

Classification and extraction

2,000,000 records/mo ≈ $112

Attributes from product copy, invoice reading, form filling, labelling. High-volume batch work.

Model: Llama 3.1 8B · Qwen2.5 7B
Context: 4K

Batch jobs get cheaper still with off-peak and reserved capacity.

Figures use Starter-tier pricing and typical request lengths. Want a number for your own workload? Tell us in the early access form and we will work it out with you.

Pricing

Three ways to buy capacity.

Token pricing for the common case, GPU-hours when you need the metal, annual contracts when procurement asks.

Token pricing

Price follows the maximum context a request may use. Long context consumes several times more GPU memory — pricing it separately keeps short-context customers from subsidising long-context ones.

Context	Starter	Growth	Typical use
4K	$0.15	$0.12	Classification, extraction, short chat
8K	$0.25	$0.20	Support assistant, RAG, summarisation
32K	$1.00	$0.80	Contract review, long documents
128K	—	$2.50	Archive search, multi-document analysis

TrialOne-time 1M tokens for new accounts. No card required, 5 req/s, no SLA.

StarterNo commitment, pay as you go. 20 req/s, shared pool.

Growth$500/month commitment, 20% off unit price. 100 req/s, 99.9% SLA, 128K context access.

Prices shown in USD. Turkish customers are invoiced in TRY with e-fatura at the contracted rate. Usage beyond the commitment continues at the same unit price.

Managed inference

$0.12 – $2.50per 1M tokens · by context

Pay per token on the shared pool. Price follows the context tier — see the table above.

OpenAI-compatible API
Streaming & function calling
50+ models
Usage dashboard & API keys
Community and email support

Join the waitlist

Most flexible

GPU rental

$0.79 – $2.50per GPU-hour

Dedicated Blackwell or Ampere GPUs for fine-tuning and custom serving. Hourly or reserved.

Bare-metal or containerised
Hourly, monthly or reserved
Private networking
Custom images
Priority scheduling

Request capacity

Enterprise

Customannual contract

Dedicated infrastructure, a signed SLA and KVKK documentation your legal team can file.

Dedicated capacity
99.9% SLA with remedies
DPA & KVKK documentation
TRY invoicing / e-fatura
Named technical contact

Talk to us

Prices shown in USD for comparability. Turkish customers are invoiced in TRY at the contracted rate.

Comparison

Where the global providers cannot follow.

Feature comparison as of publication. We keep this table honest — if a competitor ships one of these, we update the row.

Feature	LLMnow.ai	Together AI	Groq	RunPod
Turkey / MENA region	Türkiye	None	None	Europe
KVKK data residency	Full	No	No	No
Billing in TRY	Yes	No	No	No
Turkish-language support	Yes	EN only	EN only	EN only
Managed inference	Yes	Yes	Yes	Self-managed
Open-weight model choice	50+	Broad	Groq-served only	Any
Uptime SLA	99.9%	~99%	99.9%	None

Compliance

You choose the jurisdiction.

For some workloads data residency is mandatory; for others it is irrelevant. The useful question is not “are you compliant” but “under which law is my data processed”. Today our answer is Türkiye and KVKK; as regions open, each will be listed with its own legal regime.

Today: Türkiye, under KVKK Inference runs on hardware in Türkiye and requests are not forwarded to third countries. A data processing agreement drafted against KVKK is ready on paid plans.
No training on your data Customer prompts and completions are never used to train or fine-tune models. This holds regardless of the region you choose.
Zero-retention option Request a configuration where prompt and completion bodies are never written to disk. The audit trail keeps usage counters only.
Region-specific agreement The data processing agreement follows the law of the region you pick. For a controller established in Türkiye, choosing a region outside the country is a cross-border transfer — we work through that with you before signing.

Workloads outside KVKK scope — products serving customers abroad, anonymised data, public content — can run in the Türkiye region today just as well. Data residency is an option we offer, not a requirement we impose.

Early access

Get an API key before everyone else.

Tell us what you are building. We invite in batches and prioritise teams with a concrete workload.

Full name *

Work email *

Company

Role

Expected monthly volume

What are you building?

I agree that LLMnow.ai may contact me about early access and store the details above for this purpose. *

We use these details only to contact you about LLMnow.ai. See our privacy notice. Privacy notice

FAQ

Straight answers.

Is the service live today?

Our cluster in Türkiye is being brought up and the API is in private beta. Waitlist members get keys in batches as capacity lands.

Do I have to rewrite my application?

No. The API follows the OpenAI schema. Change the base URL and the API key; everything else — streaming, tool calls, embeddings — keeps working.

Where exactly is my data processed?

On GPUs in a data centre in Turkey. We do not forward inference traffic to providers outside the country.

Can I run my own fine-tuned model?

Yes, on dedicated capacity. Send us the weights and the serving configuration and we will host it.

How does TRY billing work?

Turkish entities are invoiced in Turkish Lira with e-fatura, at the rate agreed in your contract. International customers can stay on USD.

Who is behind LLMnow.ai?

A team with a background in GPU infrastructure and ML serving, operating out of Istanbul. Full details go up with the public launch.

GPU-powered LLM infrastructure for Turkey & MENA.