Use the terminal to talk to Nemotron first. Move to a server when another program needs to send requests. Add the Deep Agents client when you want Nemotron to request a file operation and receive the result.
Talk to the model in the terminal
MODEL=hf://bartowski/nvidia_NVIDIA-Nemotron-Nano-9B-v2-GGUF/nvidia_NVIDIA-Nemotron-Nano-9B-v2-Q4_K_M.gguf
version/v8/scripts/cks-v8-run run "$MODEL" \
--chat-template auto --max-tokens 64 --temperature 0Type a question at the prompt. Type /exit to close the session. To run one prompt and return to your shell, append --prompt 'Give me a short example.'. For a smaller download, start with Gemma or Qwen.
Serve the tools variant
Use a machine with memory for the 9B weights and the chosen context. The recorded Deep Agents task used a 20,480-token capacity. From the same shell and checkout, build a separate tools bundle. Allow time for compilation.
.venv/bin/python -m pip install -r server/requirements.txt
RUN="$HOME/cke-runs/nemotron-20k-tools"
version/v8/scripts/cks-v8-run serve "$MODEL" --run "$RUN" \
--serving-variant tools --context-len 20480 \
--host 127.0.0.1 --port 8080 --model-name nemotron-localFrom a second terminal run the following checks. The model ID should be nemotron-local. Check the loaded protocol against nemotron_toolcall_v1. On later launches reuse the same directory with --no-build and omit the build-time --serving-variant.
API=http://127.0.0.1:8080/v1
curl -fsS "$API/health"
curl -fsS "$API/models"
curl -fsS "$API/cke/loaded-identity"
curl -fsS "$API/responses" -H 'Content-Type: application/json' \
-d '{"model":"nemotron-local","input":"Say hello.","max_output_tokens":64}'Run the Deep Agents fixture
The harness gives the model a small workspace, a file reader, an editor and a validator. It records the tool calls and checks the final answer. Use the dependencies from the recorded trial in a client environment. Set the two hash variables to the expected serving and native session identities for your own build. Compare them with the identity response above. The client fixture creates a fresh directory for each run.
python3 -m venv "$HOME/.venvs/cke-deepagents"
CLIENT="$HOME/.venvs/cke-deepagents/bin/python"
"$CLIENT" -m pip install 'deepagents==0.7.32' 'langchain-openai==1.7.2'
"$CLIENT" server/deepagents_edit_acceptance.py \
--endpoint "$API" --model nemotron-local \
--expected-serving-identity "${SERVING_SHA:?set the expected serving identity}" \
--expected-session-library-sha256 "${SESSION_SHA:?set the expected native library hash}" \
--run-dir "$(mktemp -d)/fixture"Read report.json in the printed fixture directory. Check the tool sequence, validator result and identity before and after the run. To exercise recovery, repeat with --inject-first-read-error and another new directory.
The retained trial completed five turns using 3,242–3,519 input tokens per turn in about 20 minutes. A separate trial recovered from a failed read. These are dated results for an identified bundle. Repeated long runs, a full 20K input, and cancellation still need separate tests. Recorded Nemotron run · Harness quickstart