OpenAI-compatible inference
Use the models you need without rewriting your application.
GLM-5.3 Flash, DeepSeek V4.1 Flash, and Kimi K3 are live on the public API. Speech, OCR, document-layout, and voice models use the same account.
- OpenAI-compatible API
- 1M-token context
- Tool calling
- No training on prompts or outputs
from openai import OpenAI
client = OpenAI(
base_url="https://tarka.rest/v1",
api_key="tk_live_xxx",
)
response = client.chat.completions.create(
model="<model-id>",
messages=[{"role": "user", "content": "Hello"}],
)
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://tarka.rest/v1",
apiKey: "tk_live_xxx",
});
const response = await client.chat.completions.create({
model: "<model-id>",
messages: [{ role: "user", content: "Hello" }],
});
curl https://tarka.rest/v1/chat/completions \
-H "Authorization: Bearer tk_live_xxx" \
-H "Content-Type: application/json" \
-d '{"model": "<model-id>",
"messages": [{"role": "user", "content": "Hello"}]}'
OPENAI_BASE_URL=https://tarka.rest/v1
OPENAI_API_KEY=tk_live_xxx
Language models for agents and applications.
Compare model capabilities below. Status shows what the public API serves today.
| Model | Context | Parameters | Model IDs | Status |
|---|---|---|---|---|
| GLM-5.3 Flash Z.ai's efficient open model for high-volume chat and extraction. | 1M | 320B, 18B active | glm-5.3-flashglm-5.3-flash-eu | Available |
| DeepSeek V4.1 Flash DeepSeek's fast open model for everyday chat, tool use, and coding. | 1M | 552B backbone | deepseek-v4.1-flashdeepseek-v4.1-flash-eu | Available |
| Kimi K3 Moonshot AI's flagship open model, for coding agents and long, multi-step tasks. | 1M | 2.8T, 104B active | kimi-k3 | Available |
| GLM 5.3 GLM 5.3 through Tarka's OpenAI-compatible API. | 1M | — | glm-5.3 | Available |
| Qwen 3.8 Flash Next A free, self-hosted language model through Tarka's API. | 256K | — | qwen3.8-flash-next | Available |
Speech, documents and voice.
Transcribe audio, read documents and generate speech through the same account and API family as your language models. Open a model to see what it returns and when to pick it.
| Type | Model | Languages | License | Model IDs | Status |
|---|---|---|---|---|---|
| Speech to text | Whisper Large v3 Multilingual speech-to-text — 99 languages, including Nepali. | 99 languages | MIT (OpenAI Whisper open release) | whisper-large-v3 | Available |
| Speech to text | Qwen3-ASR Apache 2.0 speech recognition with Qwen3-ASR 1.7B. | 30 languages | Apache 2.0 | qwen3-asr | Available |
| OCR | GLM-OCR Document vision for receipts, PDFs, screenshots, forms. | General multilingual + Nepali fine-tune | MIT base; fine-tune not declared | glm-ocrglm-ocr-nepali | Available |
| OCR | PaddleOCR-VL 1.6 Compact multilingual document OCR for text, tables, formulas, charts, and seals. | Multilingual | Apache 2.0 (PaddlePaddle) | paddleocr-vl-1.6 | Available |
| Document layout | PP-DocLayoutV3 Layout analysis for text, titles, tables, formulas, figures, charts, and curved pages. | Language-independent geometry | Apache 2.0 (PaddlePaddle) | pp-doclayout-v3 | Available |
| Text to speech | Qwen3-TTS Apache 2.0 text to speech with consent-backed voice cloning. | 10 languages | Apache 2.0 | qwen3-tts | Available |
| Text to speech | Kokoro Small, fast TTS — built for low-latency agent loops. | English on the current alias | Apache 2.0 | kokoro | Available |
| Text to speech | Piper CPU-friendly English TTS with a pinned, predictable voice. | English on the current alias | GPL-3.0 runtime; voice licenses vary | piper | Available |
Keep one integration as your model needs change
Use the OpenAI SDK or any compatible HTTP client you already have. Change the base URL, pick a model ID, and switch models later without rewriting your application.
What you get
- OpenAI-compatible chat endpoints for SDKs and HTTP clients that already speak the OpenAI API format, with streaming and tool calls.
- Per-request usage records expose model, route, and token totals through the console so teams can audit workload behavior.
- No training on your prompts or outputs.
Compare partner models
Three partner models are live on the public API.
- GLM-5.3 Flash for high-volume chat, extraction, and classification.
- DeepSeek V4.1 Flash for fast, general-purpose chat, tool use, and coding.
- Kimi K3 for coding agents and long, multi-step tasks.
Open a model card to review its context window, parameter count, and model ID before
you integrate it. Only models returned by GET /v1/models for your key can be called.
Create an account and call a model.
Sign up, then request the products you need from the console.