Skip to main content

Welcome to ColabHive

Build and operate private agentic systems across heterogeneous CPU and GPU infrastructure.

ColabHive is a policy-governed control plane. It manages the CPU and GPU nodes you own as one execution layer, keeps workloads inside your cluster by default, and routes each task to the processor and node that fits it — extending into isolated capacity only when policy allows.

→ Private Agentic Infrastructure — the whole model in one page: capacity tiers, CPU/GPU execution lanes, placement, privacy, and what is available today versus in development.

→ Trust & Operations — current security controls, privacy boundaries, operational status and the explicit absence of a general public SLA.

On top of that infrastructure, ColabHive gives you a curated catalog of ready-to-run models and lets you import any compatible model from Hugging Face on demand. Both live in the same catalog and are served through the same inference path — so you are never limited to a fixed list.

  • A curated base catalog — LLMs (GPU and CPU), embeddings/rerank/translate/STT/OCR/moderation specialists, image/audio/video/speech generators, plus trainable templates (tabular, time-series, fine-tuning). Browse it live at GET /api/builder/v1/inference/models.
  • Import from Hugging Face — search a repo, check compatibility, register it, and call it. The platform builds the endpoint for you; the model downloads on its first request.

Two ways to start​

Pick the path that matches what you want to do first. You can mix them freely later.

Path 1 — Run inference now​

Call a model that is already in the catalog. No training, no setup beyond an API key.

from colabhive import ColabHive

client = ColabHive(api_key="hive_...", account_id="...")

result = client.endpoints.infer(
    "qwen-2.5-7b-instruct-public",
    {"messages": [{"role": "user", "content": "Say hello in one word."}]},
)
print(result["result"])

→ Quickstart: Inference — your first request to a real public LLM in ~2 minutes (SDK, REST, and the OpenAI-compatible endpoint).

Path 2 — Train your own​

Upload a dataset and train a model, then serve it through the same inference path.

dataset = client.datasets.upload(name="my_data", file="./train.csv")
job = client.training.create(model="xgboost-regression", dataset_id=dataset.id)
job.wait()

→ Quickstart: Training — train a model end to end.

Bring a model from Hugging Face​

Somewhere between the two: take any compatible Hugging Face repo and make it a live endpoint.

→ Quickstart: Import from Hugging Face — search → info → register → infer.

Install​

pip install colabhive

Get an API key and account ID from console.colabhive.com (Settings → API Keys). API keys start with hive_. Set them as environment variables so you don't paste secrets into code:

export COLABHIVE_API_KEY="hive_..."
export COLABHIVE_ACCOUNT_ID="..."
import os
from colabhive import ColabHive

client = ColabHive(
    api_key=os.getenv("COLABHIVE_API_KEY"),
    account_id=os.getenv("COLABHIVE_ACCOUNT_ID"),
)

The base URL defaults to https://api.colabhive.com. The Builder REST API is under /api/builder/v1; an OpenAI-compatible surface is mounted at /v1.

Outbound webhooks are a private-preview candidate

The candidate contract can send terminal notifications for inference, agent invocations, training, and merges. It remains feature-gated and absent from the default public OpenAPI; polling continues to work everywhere. See Outbound Webhooks for the exact contract and limits.

Understand the platform​

  • Private Agentic Infrastructure — the control plane, the three capacity tiers, the CPU and GPU execution lanes, and how privacy and placement actually work.
  • Platform Overview — how a request flows from the gateway through the orchestrator to GPU/CPU nodes, and the model tiers.
  • Model Catalog & Hugging Face — the difference between curated base models and Hugging Face imports, the hf-* naming, and the model lifecycle.
  • Inference Lifecycle — cold start, warm models, and how to keep latency predictable.
  • The Model Flywheel — merge and retrain models as first-class, reusable operations.
  • Elastic Cloud Burst — how the hive rents overflow GPU capacity in the cloud as a last resort, and releases it the moment it can.

Not sure which model to pick? See the Choosing a Model guide.

Documentation sections​

  • Get Started — quickstarts for inference, training, and importing from Hugging Face
  • Concepts — how ColabHive works
  • Models — the live catalog and per-model reference
  • Guides — task-oriented how-tos
  • API Reference — REST endpoints and authentication
  • SDK Reference — the colabhive Python client
  • MCP — use ColabHive from agent tooling
  • Examples — end-to-end walkthroughs

Need help?​

  • Discord: discord.gg/colabhive
  • Email: support@colabhive.com
  • Issues: github.com/colabhive/colabhive-api/issues