Choosing a model · LeSysBot Docs

The model is what decides which tool to use when you ask a question in words. A bigger one gets that right more often — it's the single biggest lever on how good LeSysBot feels. This page helps you pick one that fits your hardware.

Every model here supports tool calling, which LeSysBot needs. Using OpenAI or another remote service instead? You can skip this page — see Settings.

In a hurry: ollama pull qwen3.5:4b works on almost anything and is what the quick start uses. Move up a tier when you notice it picking the wrong tool.

1. Quick pick by GPU VRAM#

VRAM numbers are approximate at Q4_K_M quantization; add ~1–2 GB overhead for KV cache and OS.

GPU VRAM	Recommended model	Pull command
4 GB	Qwen3.5 4B (quick-start default)	`ollama pull qwen3.5:4b`
8 GB	Qwen3.5 9B ⭐	`ollama pull qwen3.5`
12 GB	Gemma4 12B	`ollama pull gemma4:12b`
16 GB	Qwen3 14B	`ollama pull qwen3:14b`
24 GB	Qwen3.5 27B	`ollama pull qwen3.5:27b`
48 GB+	Llama 3.3 70B	`ollama pull llama3.3:70b`

⭐ Best balance of quality and speed for most hardware.

Alternatives worth knowing at each tier:

Tier	Model	VRAM	Why pick it
4 GB	Llama 3.2 3B (`llama3.2`)	~3.6 GB	Smaller still (~2 GB download); weaker tool calling
8 GB	Gemma4 12B (`gemma4:12b`)	~7.6 GB	Native function calling; multimodal
12 GB	DeepSeek R1 14B (`deepseek-r1:14b`)	~9.5 GB	Strong reasoning / chain-of-thought
16 GB	Gemma4 e4b (`gemma4:e4b`)	~9.6 GB	MoE — very fast for its quality
24 GB	Qwen3 32B (`qwen3:32b`)	~22 GB	Excellent instruction following
48 GB+	Qwen3.5 122B (`qwen3.5:122b`)	~81 GB	Needs 2× 48 GB GPUs or large unified memory

Coding-heavy tools? If your LeSysBot tools involve writing or reviewing code, qwen3-coder:30b (~18 GB, MoE) is the strongest open coding model; deepseek-coder-v2 (~9 GB) is a good mid-size option.

Apple Silicon shares RAM and VRAM in one unified pool — use total RAM as your budget with the tables above (16 GB RAM → the 8–12 GB tier). Pull the -mlx tag variants for better performance: ollama pull qwen3.5:9b-mlx.

2. Tips#

Tool calling quality: Qwen3.5 and Gemma4 have the most reliable tool calling in 2026 — they rarely hallucinate function names or drop required parameters. If LeSysBot picks the wrong tool or none at all, try one of these before debugging anything else.
MoE models (e.g. Qwen3-Coder 30B) load fewer active parameters per token — faster than their total size suggests.
Quantization: Ollama picks a sensible quantization automatically. For explicit control, append :q4_K_M (smaller/faster) or :q8_0 (higher quality, more VRAM).
Out of memory: try the next smaller size or a MoE variant.
Large contexts: Qwen3.5 models have a 256K context window — you can raise agent.max_history (default 50) safely on them.

3. Managing models with Ollama#

The commands you'll actually need:

bash

ollama list                 # what's downloaded
ollama pull qwen3.5         # download a model
ollama rm qwen3.5:4b        # delete one
ollama ps                   # what's loaded in GPU memory right now
ollama stop qwen3.5         # unload a model, free VRAM
ollama run qwen3.5          # chat REPL — quick test before wiring into LeSysBot

Is the server running?

bash

curl http://localhost:11434/
# → Ollama is running

If not, start it: the official Linux installer registers a systemd service (sudo systemctl start ollama); on macOS/Windows launch the Ollama app; or run ollama serve in a terminal. For server administration beyond this — bind address, storage location, keep-alive tuning — see the official Ollama docs.

4. Pointing LeSysBot at the model#

After pulling a model, set it in config.yaml (~/.lesysbot/config.yaml for an installed setup):

yaml

llm:
  base_url: "http://localhost:11434/v1"
  model: "qwen3.5"        # ← name from `ollama list`
  api_key: "ollama"

Or one-off on the command line:

bash

lesysbot chat --model qwen3.5

Every backend (Ollama, vLLM, LlamaCpp, OpenAI) uses the same three settings — the full matrix is in Settings.