Skip to main content

Self-hosted Ollama

The B4M CLI can connect to any Ollama endpoint via --ollama-host. Models from that endpoint appear in the model picker alongside your B4M cloud models, and you can switch between them mid-session.

"Self-hosted" here means you point the CLI at a specific Ollama host — it can be:

  • Your machine — for privacy or offline use
  • A remote server or EC2 instance — shared inference across a team
  • B4M-hosted Ollama — when B4M's own API surfaces Ollama models

Connecting to an Ollama endpoint​

b4m --ollama-host <url>

On startup you'll see:

🦙 Ollama connected: 2 model(s) added to picker

Open the model picker (/model) to see both B4M cloud models and Ollama models listed together. Select any model to switch — mid-session switching is fully supported.

Examples​

# Ollama running on your machine (default port)
b4m --ollama-host http://localhost:11434

# Ollama on a remote host
b4m --ollama-host http://192.168.1.50:11434

# Ollama with basic auth
b4m --ollama-host http://user:password@ollama.internal:11434

Running Ollama yourself​

If you want to host your own Ollama instance:

Install:

# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

Windows: Download from ollama.com/download.

Start:

ollama serve   # Linux / manual start
# macOS: starts automatically after install

Pull a model:

ollama pull qwen3.5        # recommended (9b, ~7 GB RAM)
ollama pull qwen3.5:4b     # lighter option (~3 GB RAM)
ollama pull qwen3.5:27b    # higher quality (~17 GB RAM)

Qwen3.5 is the recommended model family for use with the B4M CLI. It is specifically optimized for tool calling and reasoning, which maps well to how the CLI's ReAct agent uses tools.

Limitations​

  • Tool calling performance — The CLI passes tool schemas to Ollama natively, which can add significant input tokens (~50 K for a typical session). Larger models on slower hardware may take a minute or more to respond. Using qwen3.5:4b or qwen3.5:9b gives the best balance of speed and capability.
  • Quality — Open-source models generally produce lower-quality output than frontier cloud models for complex reasoning tasks.
  • Context window — Reported as 8 192 tokens regardless of actual model capacity. This will be improved in a future release.

See Also​