Ollama
CLI + local server (port 11434). One command to pull and run models. OpenClaw: agent.baseUrl http://localhost:11434, agent.model ollama/<name>. Best for 24/7 backup, headless BRAIN, and scripting.
Local AI Models
GPT-OSS 20B, Ollama, model families, Thunderbolt storage, and equipment. Use a local model as backup to the API.
Answer a few questions on the home page to get setup suggestions tailored to your budget and goals.
GPT-OSS 20B — OpenAI’s open-weight model for local
GPT-OSS-20B is OpenAI’s open-weight model (Apache 2.0) designed for local deployment. ~21B parameters, 3.6B active (MoE), 14GB on disk with MXFP4 quantization. Runs in 16GB RAM. 128K context. Strong reasoning, chain-of-thought, function calling, and agentic use—ideal for OpenClaw backup or primary local.
ollama run gpt-oss:20b
In OpenClaw: agent.model ollama/gpt-oss:20b. There is also GPT-OSS-120B for high-end GPUs (e.g. 80GB); for Mac mini, 20B is the sweet spot.
Local model as backup to the API
Run a primary model via API (Anthropic, OpenAI) with a local model as fallback: continuity when the API is down, privacy-sensitive work on-device, and lower cost for high-volume tasks. On the BRAIN or solo Mac, run Ollama (e.g. gpt-oss:20b or llama3.1:8b) and point OpenClaw at it when you want backup. See Architecture and Best Practices.
Runtimes: Ollama vs LM Studio
LM Studio
GUI for Windows/Mac. Good for trying models. Can expose an OpenAI-compatible server. For headless or BRAIN setups, Ollama is usually simpler.
Model families and what they’re good at
| Model / Ollama tag | Best for | Size / RAM |
|---|---|---|
| gpt-oss:20b | Reasoning, agents, function calling, CoT. OpenAI open-weight. | 14GB / 16GB+ |
| llama3.1:8b, llama3.2:8b | General chat, coding. Good default. | ~4–5GB / 8GB+ |
| qwen2.5:7b, qwen2.5:14b | Strong reasoning, long context. | 7B ~4GB, 14B ~8GB / 16GB |
| gemma2:2b, gemma2:9b, gemma2:27b | Efficient; 2B for low resource. | 2B–27B / 8GB–32GB |
| phi3:mini, phi3:medium | Small, fast. Edge and quick replies. | 3B–14B / 8GB–16GB |
| deepseek-r1, deepseek-coder | Reasoning (R1), coding. | Varies / 16GB+ |
| mistral:7b, mixtral:8x7b | Speed/quality balance; Mixtral MoE. | 7B ~4GB, 8x7B ~26GB |
Equipment: RAM, storage, Thunderbolt
- RAM (unified memory): 8GB = small models (8B Q4). 16GB = GPT-OSS 20B, 14B Q4, 8B Q8. 32GB+ = 27B–70B quantized. Main limit on Apple Silicon.
- Thunderbolt 4 / USB4: Mac mini has TB4 (40 Gb/s); M4 Pro has Thunderbolt 5 (120 Gb/s). Use a TB4/USB4 NVMe enclosure for external model storage—faster than base 256GB internal in many cases. See Hardware & Storage for DRAM vs DRAMless and enclosure picks.
- Storage capacity: Models range 4GB–40GB+ each. 256GB internal is tight; external TB4 NVMe (1TB+) gives room and speed.
- CPU/GPU: Apple Silicon (M1/M2/M4) runs Ollama via Metal. No separate GPU needed for Mac mini.