Feature 09 // Intelligence routing

Models

Pick the brain. You control which model runs which task — private local models, frontier cloud models, or a mix of both. Celestos can compare models side by side, test what your hardware can actually run, and route each role to the right level of intelligence.

Try it — Get Started Read the full guide

Use cases // What people ask for

Real requests, run end to end.

Use local model for privacy-sensitive work

Switch to a frontier cloud model for hard reasoning tasks

Set standard-tier roles to a cheaper model to save money

Create a Hive that uses elite models only for the manager

What you'll see

Model picker in top bar

Model status indicators (local ✓, cloud connected)

Settings > Models page with detailed configuration

Hive role → tier → model routing visualization

Models // Intelligence Tiers

Route by difficulty, not by habit.

Tier	Meaning	Examples
Elite	≥80 IQ — frontier reasoning, hardest thinking	Frontier cloud models
Expert	≥65 IQ — strong, capable	Mid-frontier cloud models
Standard	≥50 IQ — mid-tier, fast, most steps	Smaller models, capable local models
Basic	<50 IQ — smallest, fastest	Fast local models, edge devices

Providers // Cloud + Local

Every popular cloud model provider, one control surface.

Celestos connects to the providers people actually use. The Cloud Hub loads live model catalogs, shows which providers are ready, links straight to API key setup, and lets you open remote models in the same Chat, Hive, Studio, and Tracker surfaces as your local models.

Provider	Use case
Local (on-device)	Default. Private. No API cost.
OpenRouter, Together AI, Fireworks	Unified access to broad open-model catalogs.
OpenAI, Anthropic, Google AI	Frontier reasoning, coding, vision, and general assistants.
Groq, Cerebras, SambaNova	Ultra-fast inference and free-tier experimentation.
Hugging Face, NVIDIA NIM, Cohere	Serverless, enterprise, RAG, and structured-output workflows.
DeepSeek, Mistral AI, xAI, Perplexity	Low-cost reasoning, coding, long context, and real-time web grounding.
Kie.ai, CLōD	Aggregator and curated endpoint options for cost or availability.

Cloud Hub — Live Catalogs

Ready status shows free-tier providers and configured API keys.

Live model lists load from provider catalogs instead of stale dropdowns.

One-click handoff sends any remote model straight into Chat.

Hidden model controls let you clean up noisy catalogs.

View Cloud Hub

Tips // Getting the most out of it

Field notes from real use.

Local runs via llama.cpp — private, free, no API limits

Cloud options are configurable, pay-per-use

Set default model, API keys, and parameters in Settings > Models

Intelligence tiers (Elite / Expert / Standard / Basic) route by task difficulty

Try this now

> Use local model for privacy-sensitive work, cloud
              for hard reasoning

Get Started Free

Benchmarks // Know before you commit

Know what your machine can run before you load it.

Benchmarks turns model picking into evidence. Celestos reads your CPU, RAM, GPU, VRAM, and temperature, tests hardware fit, compares up to three models side by side, and helps you choose the model that performs best on your actual machine.

Hardware fit test

Celestos scans CPU cores, RAM, GPU, VRAM, utilization, and temperature. It tells you what fits comfortably, what will be slow, and what should move to cloud before you waste time loading it.

Head-to-head comparison

Choose up to three models and compare benchmark scores, speed, VRAM, context, parameters, and license in one table. It is the fastest way to decide which model belongs in each tier.

Reference guide built in

MMLU-Pro, HumanEval, MATH, GPQA, AIME, SWE-Bench, and Arena Hard are explained inside the app, so the numbers actually mean something when you compare models.

Route by result

Assign benchmark winners to intelligence tiers. The fastest model becomes your Standard tier. The sharpest becomes Elite. The routing table updates and every future Hive run benefits automatically.

What gets measured

Tokens/second — raw generation speed on your hardware

Time to first token — how fast it starts responding

RAM usage — how much headroom you have for other work

Context window — how much it can hold at once

Output quality score — rated by a local judge model against the task

Benchmark run · 2 models · coding task

Model           tok/s   TTFT    RAM
Qwen2.5-14B     42.1    0.8s    9.2GB
Mistral-7B      78.4    0.4s    5.1GB

Quality score
Qwen2.5-14B     91 / 100
Mistral-7B      74 / 100

Recommendation: Qwen for Elite tier
Mistral for Standard (3x faster, good enough)
          

View Benchmarks