These files are external artifacts, not source: large, quantized, gitignored. A fresh clone
must place them here (workspace/models/) before the agent or the benchmark can run.
| Role | File | Size | Quant | Base model |
|---|---|---|---|---|
| NAV (orchestrator) | Qwen3-4B-Q5_K_M.gguf |
~2.9 GB | Q5_K_M | Qwen3-4B-GGUF |
| CODER | Qwen2.5-CODER-3B-Q6_k.gguf |
~2.8 GB | Q6_K | Qwen2.5-Coder-3B-Instruct-GGUF |
The filenames are load-bearing: src/config.py (NAV_MODEL, CODER_MODEL) expects these exact
names. Fetch them from the linked upstream Qwen repositories and place them here. Both upstream
model repositories declare Apache-2.0; model files remain external artifacts and are not covered
by this repository's release process.
Verify the downloaded file digest before use and record it with the validation result:
Get-FileHash .\Qwen3-4B-Q5_K_M.gguf -Algorithm SHA256
Get-FileHash .\Qwen2.5-CODER-3B-Q6_k.gguf -Algorithm SHA256benchmarks/agent_eval.py preflights their presence and fails with a pointer to this file if
either is missing, preventing a silent measurement of an incomplete environment.