zorp is meant to be usable without a hosted API, which matters most where per-token billing is the barrier. That claim currently rests on one data point.
Verified so far (2026-08-17, zorp-agent 0.2.1, Ollama):
| Model |
Result |
qwen3.8:27b |
all 67 UAT scenarios, twice |
qwen3:4b (~2.5GB, Q4_K_M) |
one-shot 4s, tool loop 5s, multi-step chain 25s, denial safety held, full validate in 86s |
Unknown: anything else. Whether a 1B to 3B model can hold a tool-calling loop, whether models without native tool support degrade usefully or fail confusingly, and what the practical floor is.
Also unmeasured: the embedding model. validate needs one, and qwen3-embedding is 4.7GB, which is larger than the 4B chat model. A smaller embedder would cut the total footprint from roughly 7GB to roughly 3GB and change what hardware this runs on. Nobody has tested whether a small embedder is good enough for the cited-source step.
How to help: run a model, report what happened. Useful shape:
export ZORP_BASE_URL=http://localhost:11434/v1 ZORP_MODEL=<your model>
zorp-agent "read README.md with your tools and reply with its first line only"
Report the model, its size and quantization, your hardware, and whether the tool call happened at all. Failures are as useful as successes: knowing where the floor is beats a vague claim that it "runs on a laptop".
The reproduction runbook for the full sweep is at the end of docs/uat/UAT-report.md.
zorp is meant to be usable without a hosted API, which matters most where per-token billing is the barrier. That claim currently rests on one data point.
Verified so far (2026-08-17,
zorp-agent 0.2.1, Ollama):qwen3.8:27bqwen3:4b(~2.5GB, Q4_K_M)validatein 86sUnknown: anything else. Whether a 1B to 3B model can hold a tool-calling loop, whether models without native tool support degrade usefully or fail confusingly, and what the practical floor is.
Also unmeasured: the embedding model.
validateneeds one, andqwen3-embeddingis 4.7GB, which is larger than the 4B chat model. A smaller embedder would cut the total footprint from roughly 7GB to roughly 3GB and change what hardware this runs on. Nobody has tested whether a small embedder is good enough for the cited-source step.How to help: run a model, report what happened. Useful shape:
Report the model, its size and quantization, your hardware, and whether the tool call happened at all. Failures are as useful as successes: knowing where the floor is beats a vague claim that it "runs on a laptop".
The reproduction runbook for the full sweep is at the end of
docs/uat/UAT-report.md.