MiniCPM5-1B-Agentic-Tooluse-v3 is a 1B-parameter open-weight function-calling model fine-tuned from MiniCPM5-1B. Quantized to GGUF, it runs entirely offline on a CPU — no GPU, no cloud API, no data leaving your machine — and works out of the box with llama.cpp, Ollama, LM Studio, koboldcpp and text-generation-webui.

The model is trained to parse a tool schema plus a natural-language request and emit a structured, correctly-named, correctly-valued function call — the core skill behind LangChain agents, LlamaIndex pipelines, AutoGen, CrewAI, MCP tool servers, ReAct loops and home-automation assistants. On a held-out 300-example benchmark it reaches 74.67% exact-argument accuracy.

Training

Unlike most small open tool-calling models that stop at supervised fine-tuning, this one goes through a two-stage pipeline:

1. QLoRA supervised fine-tuning (SFT) on tool-calling trajectories from ToolACE.
2. GRPO reinforcement learning on top of the SFT checkpoint, specifically rewarding the two hardest parts of tool calling: choosing the right function name and getting every argument value exactly right.

Compared with GPT-4o / Claude function calling, the author notes the model trades some absolute accuracy for being free, local, private, zero cost per call, and fine-tunable.

Results

Evaluated on a held-out 300-example slice drawn from a seeded shuffle of ToolACE. The base-model column is the same model with the same prompt and no adapter. The published weights are SFT + GRPO; the SFT column is kept because every negative result below is measured against it.

Metric	v2 (previous release)	SFT retrain (pre-GRPO)	v3 = SFT + GRPO (published)
parseable — output is a well-formed call	0.9933	1.0000	1.0000
valid_name — name exists among the offered tools	0.9700	0.9867	0.9867
expected_name — name matches gold	0.9067	0.9567	0.9533
args_exact — every argument value matches gold	0.6133	0.7367	0.7467
arg_key_overlap — F1 over argument keys	0.8757	0.9422	0.9388
Mean of 5	0.8718	0.9245	0.9251

GRPO buys +0.0100 on args_exact, the metric that matters most here, and gives back 0.0034 (one test example each) on expected_name and arg_key_overlap. The author reports this trade rather than hiding it: the mean moves only +0.0006, so this is a targeted gain on the hardest metric, not a broad improvement.

Full 8-Metric Benchmark (held-out test set, n=300)

Mirroring the v2 evaluation format, with Base, v2 and v3 side by side on a single consistent harness and test slice:

Metric	Base MiniCPM5-1B	v2 (previous release)	v3 (this model)	Delta (v2 → v3)
parseable_rate	0.0133	0.9933	1.0000	+0.0067
valid_name_rate	0.0133	0.9700	0.9867	+0.0167
expected_name_rate	0.0133	0.9267	0.9533	+0.0267
args_exact_rate	0.1500	0.6533	0.7467	+0.0934
arg_key_overlap	0.0033	0.7517	0.9388	+0.1871
no_schema_copy_rate	1.0000	1.0000	0.9967	-0.0033
no_repetition_rate	0.9967	1.0000	0.3400	-0.6600
stopped_cleanly_rate	0.0000	0.1500	0.0000	-0.1500

Note: the v2 values for expected_name, args_exact and arg_key_overlap differ between the two tables in the model card; they are reproduced here as published.

What the additional metrics mean:

• no_schema_copy_rate — the model did not copy the tool schema's own field description verbatim into an argument value.
• no_repetition_rate — the completion contained no duplicated function-call block or degenerate repeated-phrase loop. This is a known weakness: the model often keeps generating filler content after the tool call completes.
• stopped_cleanly_rate — the model stopped naturally right after the completed </function> tag with no trailing tokens.

So in practice, use a parser that treats the first completed <function>...</function> block as the action boundary — do not rely on natural end-of-generation.

Available Quantizations

File	Quant	Size	Best for
MiniCPM5-1B-Agentic-Tooluse-v3.F16.gguf	F16	~2.02 GB	Maximum quality, GPU or high-RAM CPU inference
MiniCPM5-1B-Agentic-Tooluse-v3.Q8_0.gguf	Q8_0	~1.07 GB	Near-lossless, recommended default for most users
MiniCPM5-1B-Agentic-Tooluse-v3.Q4_K_M.gguf	Q4_K_M	~656 MB	Smallest and fastest — edge devices, phones, CPU-only/low-RAM machines

All three run well on CPU; a GPU just makes them faster.

Quick Start

llama.cpp:

Bash

./llama-cli -m MiniCPM5-1B-Agentic-Tooluse-v3.Q8_0.gguf -p "Your prompt with tool schema here"

llama-server (OpenAI-compatible API, works with most agent frameworks):

Bash

./llama-server -m MiniCPM5-1B-Agentic-Tooluse-v3.Q4_K_M.gguf --port 8080

Ollama:

Bash

# Create a Modelfile containing:
#   FROM ./MiniCPM5-1B-Agentic-Tooluse-v3.Q8_0.gguf
ollama create minicpm5-tooluse-v3 -f Modelfile
ollama run minicpm5-tooluse-v3

LM Studio: download one of the .gguf files above directly through the LM Studio search/download UI.

Turn Thinking Off

MiniCPM5-1B ships with a built-in <think>...</think> chat template. Always use enable_thinking=False for tool calling — with thinking on, the model spends its token budget on reasoning and may never reach a completed function call. All benchmark numbers on this page use thinking off.

Python

prompt = tokenizer.apply_chat_template(
    messages, tools=tools, add_generation_prompt=True,
    enable_thinking=False,
    tokenize=False,
)

Use Cases

• Fully local / offline / private AI agents
• Home automation and smart-home voice assistants
• Mobile, browser-extension and embedded/IoT tool-calling agents
• Cost-sensitive, high-volume backends that can't afford large-model API costs per call
• A function-calling backbone for LangChain, LlamaIndex, AutoGen, CrewAI and MCP-based agent stacks
• Hobbyist and research experiments with small-model agentic reasoning

Related Repos

Format	Repository
v3 LoRA adapter (PEFT, smallest download, fine-tune further)	MiniCPM5-1B-Agentic-Tooluse-QLoRA-v3
v3 merged full-weight FP16 (transformers / vLLM / SGLang)	MiniCPM5-1B-Agentic-Tooluse-v3-Merged-FP16
v3 GGUF quantizations (llama.cpp / Ollama / LM Studio)	MiniCPM5-1B-Agentic-Tooluse-v3-GGUF
v2 LoRA adapter	MiniCPM5-1B-Agentic-Tooluse-QLoRA-v2
v2 merged FP16	MiniCPM5-1B-Agentic-Tooluse-Merged-FP16
v2 GGUF	MiniCPM5-1B-Agentic-Tooluse-GGUF

Links

• Hugging Face: MiniCPM5-1B-Agentic-Tooluse-v3-GGUF
• Base model: openbmb/MiniCPM5-1B
• Training data: Team-ACE/ToolACE