localpilot.yaml

# LocalPilot — Model Configuration
# ===================================
# Uncomment and edit to override auto-detected defaults.
# Run `python config.py --show` to see what's auto-selected for your GPU.
# Run `python config.py --models` to see all available options.

# ── Web Agent (arXiv visual browsing) ────────────────────────
# Options: MolmoWeb-4B
# Requires: 8+ GB VRAM
#
# web_agent: MolmoWeb-4B

# ── Code Agent (experiment script generation) ────────────────
# Options (best to smallest):
#   Devstral-24B-Q8    → 25 GB VRAM  (68.0% SWE-bench)
#   Devstral-24B-Q6    → 20 GB VRAM  (67.5% SWE-bench)  ← RTX 4090 / 5090
#   Devstral-24B-Q4    → 14 GB VRAM  (66.0% SWE-bench)  ← RTX 3090 / 4080
#   Qwen-Coder-14B-Q6  → 12 GB VRAM  (37.0% SWE-bench)  ← RTX 3080
#   Qwen-Coder-14B-Q4  →  9 GB VRAM  (36.0% SWE-bench)
#   Qwen-Coder-7B-Q4   →  5 GB VRAM  (33.0% SWE-bench)  ← RTX 3060
#   Qwen-Coder-7B-CPU  →  0 GB VRAM  (33.0% SWE-bench)  ← CPU only
#
# code_agent: Devstral-24B-Q6

# ── llama-server settings ────────────────────────────────────
# llama_port: 8080
# llama_ctx_size: 8192
# gpu_layers: 99      # set to 0 for CPU-only mode

# ── Custom model directory ───────────────────────────────────
# models_dir: C:/path/to/your/models