Skip to content
Kattour
Sign inGet started
An independent AI platform

An AI platform that is actually ours

Kattour trains its own language models from scratch and serves them from infrastructure we control. No third-party model sits behind the answer.

The model is ours, not rented

Kattour's language model starts from randomly initialised weights and is trained on datasets we curate, with a tokenizer we trained ourselves. No third-party checkpoint is loaded and no external AI service is called to answer a request. That is a constraint on the architecture, not a slogan: the inference engine can only serve a Kattour checkpoint.

Own weights

A decoder-only transformer with RMSNorm, rotary embeddings, grouped-query attention and SwiGLU, trained from scratch.

Own tokenizer

A byte-level BPE vocabulary learned from our own corpora, covering Arabic, German and English with no out-of-vocabulary case.

Own infrastructure

Training, inference, storage and monitoring all run on hardware we control, with real telemetry rather than estimates.

Measured, not claimed

Token counts, latency and credit charges come from the engine itself. Where a metric is unavailable, the interface says so.

One platform, built as one product

Every component is real and connected: what the Control Center shows is what the engine is doing.

Kattour Model

The transformer architecture and its training-time configuration.

Kattour Tokenizer

Byte-level BPE with reserved control tokens and a streaming-safe decoder.

Kattour Trainer

Pretraining and fine-tuning with checkpoints that resume exactly.

Kattour Inference

Bounded-concurrency generation with a KV cache and token streaming.

Kattour Knowledge

Document ingestion and retrieval that adds knowledge without retraining.

Kattour Control Center

Live operations: GPU, CPU, memory, storage, inference and training.

Use Kattour from your own code

The API follows the widely-implemented chat-completions shape, so existing clients and editor extensions work by changing three things: base URL, API key and model name. Every token is still generated by a Kattour model on our own machines.

Read the docs
curl https://your-kattour-host/v1/chat/completions \
  -H "Authorization: Bearer kattour_sk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kattour-small",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": true
  }'