LLM API

Models

Don't hard-code a model list. The models you can call depend on your project's policies, so fetch them at runtime.

List your models

Call GET /v1/models with your key to see exactly what you can use right now. This is the authoritative list — it reflects your project's access and changes as the catalog evolves.

List available models

curl https://llm.upgreat.ai/v1/models \
  -H "Authorization: Bearer $UPGREAT_API_KEY"

Each entry has an id — pass that string as the model field on a request.

Response

{
  "object": "list",
  "data": [
    { "id": "...", "object": "model", "owned_by": "upgreat" }
  ]
}

Programmatic discovery

The OpenAI SDKs expose this as client.models.list(). Use it to populate a model picker or to validate a configured model at startup, rather than maintaining a hard-coded list that drifts from reality.

Model catalog

The models we serve, with their license. Every model runs on our own hardware in Belgium; your key may see a subset, so GET /v1/models stays the authoritative list. Pricing is set per organization — see Usage.

Model	Context	Input	License
Gemma 4 31B `gemma-4-31b` Google. Long-context general model. The default for complex tasks, long documents and agents. ReasoningToolsJSON / structured output	256K tokens	text + image	Apache-2.0 Model card
Gemma 4 12B `gemma-4-12b` Google. Compact multimodal model: text, images and audio in one model. Ask questions about an audio clip directly, without a transcription step. ReasoningToolsJSON / structured output	128K tokens	text + image + audio	Apache-2.0 Model card
Mistral Small 4 119B `mistral-small-4-119b` Mistral AI. European mixture-of-experts model (128 experts, 4 active). Strong at documents, OCR and tool use. ReasoningToolsJSON / structured output Reasoning is off by default. Send "reasoning_effort": "high" to get a reasoning field. Image input is best at reading documents and screenshots; it is weaker at precise spatial layout.	256K tokens	text + image	Apache-2.0 Model card
GLM-5.3 Flash `glm-5.3-flash` Z.ai. Natively multimodal mixture-of-experts (320B total, ~18B active) with a 1M-token context. Strongest at reasoning and agentic tool use. ReasoningToolsJSON / structured output Reasoning is on by default. Tune it with "reasoning_effort": "low", "high" (default) or "max". Many long prompts sent in parallel queue for prefill; spread large batch jobs.	1M tokens	text + image	MIT Model card
Qwen3.8 27B `qwen3.8-27b` Qwen (Alibaba). Fast dense 27B model with native vision and a 262K-token context. Strong at coding, reasoning and agentic tool use. ReasoningToolsJSON / structured output Thinking is on by default ("reasoning_effort": "xhigh"; "medium" and "low" also work, and "high" is treated as "xhigh"). Turn it off per request with "chat_template_kwargs": {"enable_thinking": false}.	256K tokens	text + image	Apache-2.0 Model card
DeepSeek V4.1 Flash `deepseek-v4.1-flash` DeepSeek. Successor of V4 Flash: fast mixture-of-experts with native vision and a 1M-token context. Built for agents, long documents and screenshots. ReasoningToolsJSON / structured output Reasoning is on by default and returned in the "reasoning" field. Tune it with "reasoning_effort": "low", "high" (default) or "max", or send "none" to turn it off. Many long prompts sent in parallel queue for prefill; spread large batch jobs.	1M tokens	text + image	MIT Model card
DeepSeek V4 Flash `deepseek-v4-flash` DeepSeek. Fast mixture-of-experts (284B total, ~13B active) with hybrid sparse attention. The high-throughput executor for agents and long documents. ReasoningToolsJSON / structured output Reasoning is on by default and returned in the "reasoning" field (not "reasoning_content"). Turn it off per request with "chat_template_kwargs": {"thinking": false}.	256K tokens	text	MIT Model card
NuExtract 3 `nuextract-3` NuMind. Small vision-language model for structured extraction: turns text, scans and page images into JSON that follows your template. Structured extractionJSON / structured output	32K tokens	text + image	Apache-2.0 Model card
PaddleOCR-VL 1.6 `paddleocr-vl-1.6` PaddlePaddle (Baidu). Document OCR: send one page image, get the text back. Also reads tables. OCR Use the chat endpoint with one page image per request and the prompt "OCR:" (or "Table Recognition:" for a table). Split PDFs into page images on your side.	16K tokens	text + image	Apache-2.0 Model card
Qwen3Guard Gen 4B `qwen3guard-gen-4b` Qwen (Alibaba). Safety classifier for prompts and model responses. Returns Safe, Unsafe or Controversial with a category. Works in Dutch and English. Moderation Call it like a chat model: send the conversation to check and read "Safety: …" and "Categories: …" from the reply (plus "Refusal: …" when the last message is from the assistant). Use "temperature": 0 and "max_tokens": 64.	16K tokens	text	Apache-2.0 Model card
Qwen3 Embedding 8B `qwen3-embedding-8b` Qwen (Alibaba). Multilingual text embeddings for search, clustering and RAG. Embeddings	8K tokens	text	Apache-2.0 Model card
BGE-M3 `bge-m3` BAAI. Compact multilingual embeddings; a fast, low-cost option for retrieval. Embeddings	8K tokens	text	MIT Model card
Qwen3 Reranker 8B `qwen3-reranker-8b` Qwen (Alibaba). Reranker for search and RAG: scores how well each document answers a query. Works across languages, including Dutch. Reranking Send "query" and "documents" to POST /v1/rerank and read "relevance_score" per document. A document longer than the context is cut to 8K tokens. Billed on input tokens only.	8K tokens	text	Apache-2.0 Model card
Whisper Large v3 `whisper-large-v3` OpenAI. Speech-to-text in 99 languages, the most accurate transcription option. Transcription	—	audio	Apache-2.0 Model card
Whisper Large v3 Turbo `whisper-large-v3-turbo` OpenAI. Faster, lower-cost speech-to-text with the same language coverage. Transcription	—	audio	MIT Model card
Qwen3 ASR 1.7B `qwen3-asr-1.7b` Qwen (Alibaba). Fast speech-to-text, including Dutch and English. Returns plain text. Transcription	—	audio	Apache-2.0 Model card

Capabilities by endpoint

Models are grouped by what they do, and each capability maps to one endpoint. Look up the concrete model IDs for each with GET /v1/models.

Capability	Endpoint
Chat & text General chat, reasoning, translation and other text generation. Reasoning models also return a reasoning field.	`/v1/chat/completions`
Structured extraction Turn unstructured documents into JSON that matches a template you provide.	`/v1/chat/completions`
Embeddings Vectors for semantic search, clustering and RAG.	`/v1/embeddings`
Reranking Order documents by how well they answer a query, for search and RAG.	`/v1/rerank`
Image generation Text-to-image and image editing. Output is base64-encoded PNG.	`/v1/images/generations`
Audio transcription Speech-to-text from an uploaded audio file.	`/v1/audio/transcriptions`

Video (preview)

Image-to-video and speech-to-video generation are available on request — talk to your account contact.