Tholos-2B GGUF: local AI agents with Ollama and llama.cpp

GGUF builds of Tholos-2B, a small language model for AI agents that choose tools for tables, notes and tasks. Tholos-2B passes 137/160 scenarios; MiniCPM5-2B passes 112/160 on Tholos-Bench with Q4_K_M, llama.cpp JSON schema decoding and one Kaggle T4. Only Tholos-2B was fine-tuned for this task format.

Run with Ollama

ollama pull hf.co/mertkayacs/Tholos-2B-GGUF:Q4_K_M && uv tool install git+https://github.com/mertkayacs/tholos && tholos

The same model is also in the Ollama library as mertkayacs/tholos-2b (ollama pull mertkayacs/tholos-2b).

Open http://127.0.0.1:7070, select Detect in Settings, then add the model. The repo's template and params preserve the trained prompt format, stop tokens and 16,384-token context.

Direct Ollama clients should request json_object mode and reasoning_effort: "none", then validate each step. Tholos does this for you and applies approval rules.

Run with llama.cpp

llama-server -hf mertkayacs/Tholos-2B-GGUF:Q4_K_M --jinja -c 16384 -t 4 -a tholos-2b --host 127.0.0.1 --port 8080

Use json_schema response format. The complete request shows the tool definitions, step schema and response loop.

Files

Use Q4_K_M to match the benchmark. SHA256SUMS records the checksums. File size is not total runtime RAM.

Results and limits

With Q4_K_M on a CPU, Tholos-2B passes 134/160 with llama.cpp JSON schema, against 117/160 for MiniCPM5-2B. With Ollama JSON mode and the repo template, the counts are 136/160 and 96/160. Compare results within one setup.

English only. The model learned Tholos's thirteen tools; other formats are unmeasured. Validate calls, treat external text as untrusted, and keep approvals enabled for consequential actions.

Apache-2.0. Full-precision weights, training and evaluation | Code | Kaggle | Project | Benchmark page.

An Eschatia Labs project. Mert Kaya.