Quick Start

Quick Start

60-second tour

git clone https://github.com/mtecnic/model-chat-cli.git
cd model-chat-cli
pip install -r requirements.txt
python main.py

That's everything. The next five things will happen, in this order:

1. Subnet scan. Auto-detects your /24 and probes ports 11434, 1234, 5000, 8000, 8080. Discovery
2. Model table. Lists every model on every server with latency. Discovery#discovered-models-table
3. Picker. Arrow keys, Enter to chat. Type a /command from the chat prompt for everything else.
4. Cache write. Server list goes to ~/.model_chat_cache.json so next launch is instant.
5. You're chatting. Streaming response with decode TPS / TTFT in the status line. Chat

---

The five commands you'll use

  /stress       → 6 modes of load testing                 [[Stress-Testing]]
  /arena        → Multi-model head-to-head                 [[Arena]]
  /promptarena  → System-prompt tournament                 [[Prompt-Arena]]
  /think        → Toggle reasoning-token rendering         [[Chat#thinking-mode]]
  /export       → Dump conversation to markdown

Full command list: Chat#commands.

---

Common first-run gotchas

Symptom	Fix
"No servers found"	Your model host isn't on the same /24. Check firewall / VLAN. Troubleshooting#no-servers-found
Latency column shows huge numbers	Often a cold-cache LM Studio model — first request loads weights. Re-run scan.
/think shows nothing	Model doesn't emit <think> tags. Try a Qwen 3 / 3.5 reasoning variant.
Stress test exits immediately	The selected server's model name has changed. Re-pick from /switch.
Arena tournament hangs at "Judging…"	Judge model failed mid-response; check logs/stress_test_*.log for the trace.

---

What to try next

• Run a Throughput stress test at concurrency 20 to see how your server scales: Stress-Testing#throughput.
• Try Tool Bench → Quick against any model that advertises tool-calling: Tool-Calling-Benchmark#quick-suite.
• Run an Arena → Quick Compare between two models you're choosing between: Arena#quick-compare.