Model Catalog

ColabHive is not a fixed list of models. The catalog is two things at once:

1. Import any model from HuggingFace. Search HuggingFace, check compatibility, and register a repo — ColabHive creates the model config and a servable endpoint for it (cold-started on first request). See Import from HuggingFace.
2. A curated base catalog maintained by ColabHive — ready-to-call LLMs, specialists, tools, generative models, and trainable templates.

Because the catalog is data-driven, this page does not enumerate every model or repeat volatile numbers (context windows, VRAM, prices, throughput). For an authoritative, always-current list, query the live catalog:

# Public inference catalog (no auth) — grouped by category, with live readiness
curl "https://api.colabhive.com/api/builder/v1/inference/models?include_readiness=true"

# Public endpoints you can call right now
curl "https://api.colabhive.com/api/builder/v1/endpoints?visibility=public"

from colabhive import ColabHive

client = ColabHive(api_key="hive_...", account_id="YOUR_ACCOUNT_ID")
for ep in client.endpoints.list(visibility="public"):
    print(ep)

Readiness is dynamic

A model's serving state (warm / cached / cold) changes with usage — popular models stay warm automatically. Pass include_readiness=true to see current state before you call. See the Inference Lifecycle.

---

Curated base vs. imported (hf-*)

	Curated base	Imported from HuggingFace
Name	plain (e.g. qwen-2.5-7b-instruct, specialist-embeddings)	prefixed hf- (e.g. hf-google-gemma-2-9b-it)
How it got here	published by ColabHive	you registered it via POST /models/hf/register
Price	some are paid (per request, USD), some free	currently free (price_per_request = 0)
Visibility	curated / public	forced public + base on registration (see note)

HuggingFace imports are public, free, and "base" today

When you register a model through the HuggingFace path, the platform currently records it as visibility=public, lifecycle_status=candidate, is_base_model=true, and price_per_request=0. The schema rejects non-public or non-candidate registration values with HTTP 422 instead of ignoring them. (Trained-model registration via register_for_inference is different: it supports account/public with review and pricing.)

Names can differ from the docs

The doc label and the real model_name sometimes diverge (e.g. this doc's "Gemma 2 9B" is hf-google-gemma-2-9b-it in the system; llama-3.2-8b-instruct is actually Llama 3.1 8B). Always use the real model_name / endpoint name from the live catalog.

---

Inference tiers

How a model is served defines its tier. All tiers are reached through the same Inference API; the OpenAI-compatible /v1/chat/completions surface is also available for chat models.

🤖 LLM Inference (GPU)

Pre-trained chat / code / completion LLMs served on GPU via vLLM or transformers, plus a multimodal vision-language model. Includes curated base models and HuggingFace imports. → LLM Inference

🖥️ CPU LLM Inference (llama.cpp / GGUF)

Run LLMs entirely on CPU and system RAM — no GPU — over GGUF quantized weights (inference_engine=llamacpp). Wired into the runtime and served through the normal path; expect a few tokens/second. There are no active GGUF models in the live catalog today — the tier activates when you register a GGUF repo. → CPU LLM Inference

🎯 Specialists

Task-specific endpoints: embeddings, reranking, translation, OCR, speech-to-text, and moderation. → Specialists

🎨 Generative

Image, audio/voice, and video generation, served through the diffusers/generative backends. → Generative Models

🔁 Merged & Retrained

Models you build by merging adapters into a base or retraining on top of one appear here as ordinary model_versions — servable through the normal path and reusable as the base for the next round. This is the model flywheel. → Merged & Retrained Models

🔧 Tools

Network-enabled utilities (web fetch/search/scrape, geocoding) callable as endpoints or MCP actions. → Tools & Utilities

---

Trainable model templates

Beyond inference, ColabHive ships templates you can fine-tune on your own data (they carry a script_template). These are training entry points, not ready-to-call inference endpoints.

• ML Classical (tabular): gradient boosting (XGBoost, LightGBM, CatBoost), Random Forest, linear / logistic regression, SVM / SVR, NGBoost. Several support multi-output regression. Some have a GPU-accelerated variant (a distinct model_name, e.g. xgboost-regression-gpu); the CPU and GPU variants are separate configs — check the catalog for the exact name.
• Deep Learning (tabular): MLP, TabNet.
• NLP training: BERT classification; LLM fine-tuning via QLoRA / PEFT; and SSM/hybrid fine-tuning templates (mamba-ssm-finetune, jamba-hybrid-finetune). SSM/hybrid architectures are available for fine-tuning only — they are not inference endpoints.
• Time Series: forecasting (ARIMA, Prophet, PatchTST, BiTCN, ACEformer, TimesFM 2.5, multivariate, auto-selection) and a Volatility & Regime family (GARCH, HAR-RV, HMM).

See Choosing a Model for selection heuristics, and list trainable configs programmatically:

for c in client.training.model_configs():
    print(c.model_id, c.display_name)

---

Import any HuggingFace model

The registry supports importing arbitrary HuggingFace repositories:

# 1) Search (only_supported filters to architectures ColabHive can serve)
results = client.models.hf.search(query="mistral", task_type="text-generation", max_size_gb=20)
for m in results:
    print(m.repo_id, m.compatibility.status)   # dataclasses — attribute access

# 2) Inspect one (compatibility + estimated requirements)
details = client.models.hf.info("mistralai/Mistral-7B-Instruct-v0.3")

# 3) Register (task_type is REQUIRED) — creates model_config + endpoint
res = client.models.hf.register(
    repo_id="mistralai/Mistral-7B-Instruct-v0.3",
    task_type="text-generation",
)
print(res.endpoint_id)   # usable; cold-starts on first request

task_type uses the ColabHive vocabulary: embeddings, rerank, translation, ocr, stt, moderation, text-generation. See the Models API for the full contract and the Import from HuggingFace guide for an end-to-end walkthrough.

---

Learn more

• Quick Start — end-to-end workflow
• Choosing a Model — selection heuristics by task, size, hardware
• Compatibility Matrix — supported architectures and backends
• Inference API · Models API

---

Authors: José Luis Minich, Maximiliano Lucius.