Hardware-aware model selection

Probe your CPU, RAM, GPU, and disk to recommend per-task LLM models — then optionally apply them to ai-skill-model profiles.

The model advisor probes your machine and recommends per-task LLM models for Arka’s skill profiles (route, chat, summarize, etc.).

Natural language

arka select best model for my pc
arka what llm model should I use
arka optimize models for my hardware
arka best models for my mac
arka auto configure models from my hardware --apply

CLI

select_model                    # show recommendations
select_model --apply            # write to llm-skill-models.json
select_model --json             # machine-readable output
select_model probe              # hardware snapshot only
select_model --local            # choose the strongest installed local model that fits
select_model --local --run "summarize this repo"  # run a prompt with it via Ollama

Hardware tiers

Tier	Typical hardware	Strategy
minimal	Low RAM, no GPU	Cloud-only, tiny/fast models
cloud_light	Laptop, no local LLM	Fast cloud APIs
balanced	8–16 GB RAM	Cloud primary, optional small local
local_capable	16+ GB RAM, GPU	Mix Ollama + cloud for heavy tasks
local_heavy	32+ GB RAM, strong GPU	Prefer local Ollama models

The advisor checks CPU cores, RAM, GPU (CUDA/MPS/Metal), disk space, battery status, and installed Ollama models.

Apply recommendations

select_model --apply
# or
arka auto apply models for my hardware

This writes to ~/.config/arka/llm-skill-models.json. Override individual profiles anytime:

arka ai-skill-model chat gemini/gemini-2.0-flash

Requirements

• At least one cloud API key (GEMINI_API_KEY, GROQ_API_KEY, etc.) for cloud tiers.
• ollama serve running with pulled models for local tiers.

For an on-device-only choice, use select_model --local. Arka inspects the available Ollama models and local memory/GPU capacity, then selects the largest installed text model that should fit with runtime headroom. Add --apply to use that model for all skill profiles.

Related

• LLM concepts — providers and failover
• Configuration reference — ai-skill-model settings