MiniCPM5-1B Agentic Tooluse Nemotron DPO GGUF

Newly Updated and Better Checkpoint (Click Here)

Current GGUF exports of the July 2026 Nemotron SFT+DPO repaired MiniCPM5-1B-Agentic-Tooluse-Merged-FP16.

These files were generated from the latest evaluated merged checkpoint. The previous warning that this repository had not been regenerated is no longer applicable.

Model Family

Current Files

The pre-Nemotron exports are retained under legacy/ for reproducibility. Use the root-level Nemotron-DPO files for the current model.

Provenance

• Base model: openbmb/MiniCPM5-1B
• Current adapter: ewinregirgojr/MiniCPM5-1B-Agentic-Tooluse-QLoRA-v2
• Conversion source: ewinregirgojr/MiniCPM5-1B-Agentic-Tooluse-Merged-FP16
• Source revision: 67e1cad9b15ffc21d0877bf952d9ece2eeda5d35
• llama.cpp revision: 683f0c72e5b3c07fab90bfd9ec2ce8661d624228
• Quantizations: F16, Q8_0, Q4_K_M

The export pipeline checked each output's GGUF magic, format version, size, and successful completion by llama-quantize. Full task evaluation was performed on the source merged model, not independently on every quantization.

Tool-Call Format

The model emits XML-style calls:

<function name="tool_name"><param name="parameter">value</param></function>

A deployment runtime should:

1. Provide the available tool definitions in the training-compatible prompt format.
2. Decode deterministically for tool selection.
3. Stop after the first complete </function>.
4. Validate the function name and arguments against the supplied schema.
5. Execute the tool outside the model and provide the result in a new turn.

llama.cpp

Use a recent llama.cpp build:

llama-cli \
  -m MiniCPM5-1B-Agentic-Tooluse-Nemotron-DPO.Q4_K_M.gguf \
  -p '<user>Fix the calculator bug.</user>
<tools>{"name":"run_tests","description":"Run the test suite."}</tools>
<calls>' \
  -n 96 \
  --temp 0

The exact chat template and EOS handling matter. If a frontend supplies its own template, verify that it matches the source model's chat_template.jinja.

vLLM

As of July 2026, vLLM documents GGUF support as highly experimental and under-optimized. GGUF support also requires the out-of-tree plugin:

uv pip install vllm vllm-gguf-plugin

Serve Q4_K_M directly from the Hub:

vllm serve \
  ewinregirgojr/MiniCPM5-1B-Agentic-Tooluse-GGUF:Q4_K_M \
  --tokenizer ewinregirgojr/MiniCPM5-1B-Agentic-Tooluse-Merged-FP16 \
  --hf-config-path ewinregirgojr/MiniCPM5-1B-Agentic-Tooluse-Merged-FP16

The separate tokenizer/config arguments follow vLLM's recommendation to avoid unstable GGUF tokenizer conversion. The merged repository is used because it contains the matching tokenizer, configuration, and chat template.

For stable high-throughput vLLM deployment, prefer the merged safetensors repository when memory permits.

Source-Model Evaluation

External evaluation used 300 examples derived from the untouched Team-ACE/ToolACE dataset. The same cases and deterministic first-call evaluator were used for the base and fine-tuned source model. This is not an official ToolACE leaderboard submission.

These scores establish the behavior of the merged source checkpoint. Quantization can alter individual outputs, so evaluate the exact GGUF and runtime configuration used in production.

Source-Model Improvements and GGUF Scope

These GGUF files were converted from the repaired merged checkpoint, whose reported results improved substantially over base MiniCPM5-1B:

• Parseable calls: 1.33% -> 99.33%
• Valid available-tool names: 1.33% -> 97.00%
• Expected-tool selection: 1.33% -> 92.67%
• Exact arguments: 15.00% -> 65.33%
• Argument-key overlap: 0.33% -> 75.17%
• No repetition: 99.67% -> 100.00%
• Natural clean termination: 0.00% -> 15.00%

These are source FP16 metrics. The remaining gap to 100% is residual error after improvement, not evidence that the fine-tune made the base model worse.

Deployment Notes

• Q4_K_M and Q8_0 have not yet received an independent 300-case ToolACE run, so their exact scores should not be assumed identical to FP16.
• Schema validation and permission checks are universal tool-runtime safety requirements, not model-specific regressions.
• Valid-name and exact-argument behavior improved strongly in the source model; rare errors can still occur on unseen tools.
• MiniCPM5's parser-based serving path extracts the first completed XML call instead of relying solely on natural EOS.
• vLLM GGUF support is experimental; merged safetensors remains the more reliable vLLM format.
• llama.cpp behavior depends on build version and the frontend's chat-template handling.

Version History

The current Nemotron-DPO files are the July 2026 repaired exports. Legacy pre-Nemotron files are stored under legacy/ for reproducibility, and earlier repository states remain available from Hugging Face commit history.

MiniCPM5 Tool-Call Runtime Contract

MiniCPM5-1B has native XML-style tool calling. OpenBMB recommends SGLang for tool calling with its built-in minicpm5 parser:

python -m sglang.launch_server \
  --model-path MODEL_PATH \
  --tool-call-parser minicpm5

The parser converts a completed <function ...>...</function> block into an OpenAI-compatible tool_calls response. A production agent should treat the first completed function block as the action boundary, validate it, execute the tool externally, and send the tool result in a new turn.

The reported stopped_cleanly_rate=0.15 is a strict natural-termination metric: only 15% of evaluated generations emitted a completed call and then naturally terminated at that boundary without runtime intervention. It does not mean that only 15% produced usable tool calls. On the same 300 cases, the repaired source model achieved:

• parseable_rate=0.9933
• valid_name_rate=0.9700
• expected_name_rate=0.9267
• no_repetition_rate=1.0000

Therefore, natural EOS-at-boundary remains weak, while first-call extraction and selection are much stronger. Use the official parser where supported and configure bounded generation or parser-side stopping after the first complete </function>.

Base model architecture

MiniCPM5-1B uses a standard LlamaForCausalLM architecture:

Thinking mode

MiniCPM5-1B has a built-in <think>...</think> chat template. The same checkpoint can act as a fast assistant or a deliberate chain-of-thought reasoner — controlled by a single flag:

# Fast mode — recommended for tool calling (thinking OFF)
prompt = tokenizer.apply_chat_template(
    messages, tools=tools, add_generation_prompt=True,
    enable_thinking=False,
    tokenize=False,
)

# Reasoning mode (thinking ON — NOT recommended for tool calling)
prompt = tokenizer.apply_chat_template(
    messages, tools=tools, add_generation_prompt=True,
    enable_thinking=True,
    tokenize=False,
)

Important: always use enable_thinking=False for tool/function calling. With thinking ON the model spends its token budget inside <think>...</think> and may not reach a completed function call. All benchmark numbers in this card use thinking OFF.

Citation

If you use this model, please cite the base model paper:

@article{minicpm4,
  title   = {MiniCPM4: Ultra-Efficient LLMs on End Devices},
  author  = {MiniCPM Team},
  journal = {arXiv preprint arXiv:2506.07900},
  year    = {2025}
}

And the ToolACE dataset used for fine-tuning:

@article{toolace,
  title   = {ToolACE: Winning the Points of LLM Function Calling},
  author  = {Liu, Ying and others},
  journal = {arXiv preprint arXiv:2409.00920},
  year    = {2024}
}

ModelScope

The base model is also available on ModelScope (for users in China and East Asia):

• OpenBMB/MiniCPM5-1B on ModelScope

(The fine-tuned adapter/GGUF builds are currently HuggingFace-only.)

Related repos

Latest v3 release

v2 family (this release)

Limitations