Specialists
Task-specific endpoints for embeddings, reranking, OCR, speech-to-text, and moderation.
Specialists are small, purpose-built models that do one job well — faster and cheaper than steering a
general LLM at the same task. They run on the specialist backends (GPU or CPU) and are called through
the same Inference API as every other model: POST /endpoints/{id}/infer
with an input object.
curl "https://api.colabhive.com/api/builder/v1/endpoints?visibility=public"
Prices are price_per_request in USD (the platform bills per request in USD — there is no
"HiveCredits" unit). This page does not repeat prices as fixed numbers; read them from the catalog.
Public specialist endpoints
| Task | Endpoint(s) | Model | Notes |
|---|---|---|---|
| Embeddings | embeddings-public · qwen3-embeddings | all-MiniLM-L6-v2 (384-dim) · Qwen3-Embedding-0.6B | Semantic search, RAG, clustering → details |
| Embeddings (8B) | hf-Qwen-Qwen3-Embedding-8B (free) | Qwen3-Embedding-8B | Top-tier retrieval quality → details |
| Rerank | rerank-public · qwen3-reranker | BAAI/bge-reranker-base · Qwen3-Reranker-0.6B | Reorder retrieved docs by relevance → details |
| Rerank (8B) | hf-Qwen-Qwen3-Reranker-8B (free) | Qwen3-Reranker-8B | Max-precision causal reranking → details |
| Speech-to-text | stt-public (⚠ candidate) · stt-turbo | Whisper small · Whisper large-v3-turbo | Transcription → details. Prefer stt-turbo: stt-public is still candidate in the catalog |
| OCR | ocr-public | Tesseract | Text from images, CPU-only → details |
| Moderation | moderate-public | unitary/toxic-bert | Toxicity scoring → details |
| Translation | translate-public | NLLB-200-distilled-600M | ⚠ not yet implemented → details |
There are also HuggingFace-imported embedding models in the catalog (e.g.
hf-sentence-transformers-all-MiniLM-L6-v2, hf-jinaai-jina-embeddings-v2-base-de), and you can
import any HuggingFace model for these task types
(embeddings, rerank, translation, ocr, stt, moderation).
Quick start
from colabhive import ColabHive
client = ColabHive(api_key="hive_...", account_id="YOUR_ACCOUNT_ID")
# Embeddings — pass texts, get vectors back
result = client.endpoints.infer(
endpoint_id="embeddings-public", # SDK resolves the name to a UUID
input_data={"texts": ["ColabHive is a distributed AI platform",
"Machine learning on decentralized GPUs"]},
)
print(result["result"] if result.get("status") != "queued"
else client.endpoints.get_task(result["task_id"]))
curl -X POST "https://api.colabhive.com/api/builder/v1/endpoints/{ENDPOINT_ID}/infer" \
-H "X-API-Key: hive_..." \
-H "Content-Type: application/json" \
-d '{"input": {"texts": ["hello world"]}}'
Each specialist has its own input shape (see its detail page). The authoritative request/response
schema for any endpoint is GET /api/builder/v1/endpoints/{id}.
Why a specialist instead of an LLM?
- Cheaper and faster per call — these models are small and single-purpose.
- Structured output — fixed JSON shapes (vectors, scores, labels) instead of free text to parse.
- The right tool — for search/RAG always embed + rerank rather than asking an LLM to rank.
Learn more
Authors: José Luis Minich, Maximiliano Lucius.