Ship AI productson open models,with one API.
Qwen 3.8, gpt-oss, Llama 3.3 and more behind one OpenAI-compatible endpoint. Streaming, tool calling, vision and IQD billing — four lines to integrate.
curl https://api.kalla.website/v1/chat/completions \
-H "Authorization: Bearer $KALLA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen3.8",
"stream": true,
"messages": [
{ "role": "user", "content": "Explain SSE in one line." }
]
}'- Models served
- 8
- Qwen · gpt-oss · Llama
- API surface
- 1
- OpenAI-compatible /v1
- Lines to integrate
- 4
- Change the base URL
- Billing currency
- IQD
- No FX surprises
Speaks the same protocol as the tools you already use
Eight models. One endpoint.
Qwen 3.8/3.5, gpt-oss and Llama 3.3 — from a 9B workhorse to a 397B mixture of experts. Every model speaks the same OpenAI-compatible dialect, so switching is a one-word change.
Qwen3.5-Lite
Fast, cheap 9B for chat, extraction and drafting.
gpt-oss-pro
Open-weight 120B mixture-of-experts with strong reasoning per euro.
gpt-oss-lite
Small gpt-oss for high-volume chat and classification.
Qwen2.5-VL
Vision-language model for documents, screenshots and charts.
More models, same key
The catalogue is a database row, not a deploy. New sizes, embedding models and future families land on the same endpoint — your integration never changes.
Everything a product team needs
to ship on day one.
Not a demo endpoint. Quotas, audit trails, streaming and edge-shaped errors — the unglamorous parts that decide whether an API survives production.
OpenAI-compatible, honestly
POST /v1/chat/completions with the request and response shapes your SDK already knows. Change the base URL and the key — nothing else.
Token-by-token streaming
Server-sent events end to end. First token in, not last token out.
Per-key control
Scopes, expiry, RPM and monthly caps per key. Revoke or rotate in one click — effective instantly.
Tool calling & JSON mode
Function calling and structured output for agentic workflows and reliable parsing.
Fair-use scheduler
A bounded queue absorbs bursts. At peak you wait a moment instead of getting dropped.
Built like a bank
Keys are hashed with HMAC-SHA256, never stored in plaintext. Every admin action is audited.
Busy hour? You get a queue, not a 500.
Every request passes through a bounded scheduler before it reaches the GPUs. Below capacity, responses are immediate. Above it, requests wait in line and are served in order — slower, never silently lost.
A ledger, not a guess
Every call is recorded with model, tokens, latency and status. Your team sees the same numbers we do — no invoice surprises.
- Requests this month
- per key
- Token split
- prompt / completion
- Latency
- p50 · last 30 days
- Errors
- by code
Drop it into the tools you already open daily.
If a tool speaks OpenAI, it speaks Kalla. Point the base URL at our gateway, paste your key, pick a model id.
Cursor
Point Cursor at Kalla and use Qwen for chat and inline edits.
Settings → Models → OpenAI API
Base URL https://api.kalla.website/v1
API key kalla_live_xxxxxxxxxxxxxxxx
Models Qwen3.8 # vision + tools
Qwen3.5 # flagship
Qwen3.5-Lite # fast
# Enable "Override OpenAI Base URL", add the model ids manually,
# then pick Qwen3.8 in the model selector.One flat price. Priced in dinar.
Per model, unlimited tokens within fair use. No metering per thousand tokens, no exchange-rate surprises on your invoice.
Billed monthly. Cancel any time — keys are revoked on the spot.
- 1Tell us what you are buildingOne WhatsApp message, no forms.
- 2We provision the key and limitsUsually within one business day.
- 3Point your editor at /v1Cursor, VS Code, opencode or your backend.
Message us on WhatsApp — keys are usually issued the same day.
Included with every key
- All eight models — switch per request
- OpenAI-compatible API with token streaming
- Tool calling, JSON mode and long context
- Per-key rate limits, quotas and usage ledger
- Editor setup support: Cursor, VS Code, opencode
- Support in Arabic and English
- Dedicated GPU capacity — available on request
- Fine-tuning — on the roadmap
When the shared GPUs are busy, requests wait in a short queue and responses can be slower — nothing is dropped.
Straight answers, no asterisks.
Still unsure about something? Message us on WhatsApp — a human who deploys this stack replies.
Yes. The paths (/v1/chat/completions, /v1/models), request body, response body and SSE chunk format match OpenAI. Official SDKs, Cursor, Continue, Cline and opencode work by changing only the base URL and API key.
Bring your own editor.
We'll bring the models.
Tell us what you are building and we will provision a key, tune your limits, and stay on the line while you ship.
baseURL: "https://api.kalla.website/v1"